<p>Agentic systems are moving beyond prompt driven responses into something far more structured and deliberate. They observe, reason, act, and refine. At the centre of this loop sits memory. Not as a passive store, but as an active component that shapes behaviour, context, and outcomes over time.</p>
<p>In enterprise environments, especially those dealing with operational intelligence, memory design becomes a defining factor in whether an agent behaves like a novelty or a reliable system.</p>
<div class="quote">
"Agents without memory repeat work. Agents with memory build capability."
</div>
<h2>The Layers of Memory in Agentic Systems</h2>
<p>Memory in agentic AI is not a single construct. It is a layered architecture, each type serving a distinct purpose in the decision cycle.</p>
<h3>1. Short Term Memory (Session State)</h3>
<p>This is the working context of the agent. It holds the immediate state of interaction, including recent inputs, intermediate reasoning steps, and current objectives.</p>
<p>In practical terms, this is what enables an agent to maintain coherence during a session.</p>
<ul>
<li>Maintains conversational or task continuity</li>
<li>Stores reasoning traces and tool outputs</li>
<li>Resets frequently to avoid drift</li>
<li>Closely tied to token or context window constraints</li>
</ul>
<h4>Example Use Cases</h4>
<ul>
<li>Items currently in a shopping basket before checkout</li>
<li>The last few stock prices seen in a trading app</li>
</ul>
<h3>2. Episodic (Long Term Memory)</h3>
<p>This layer captures experiences and persists them over time. It includes structured records of past interactions, decisions, and outcomes, forming a growing body of knowledge that the agent can draw upon.</p>
<p>This is where agents begin to accumulate value rather than simply respond.</p>
<ul>
<li>Stores past tasks, interactions, and outcomes as durable events</li>
<li>Enables recall of similar scenarios across sessions</li>
<li>Capture user preferences for language, report format, and notification settings</li>
<li>Supports reflection and longitudinal improvement</li>
<li>Often implemented using vector stores combined with durable storage layers</li>
</ul>
<div class="pullquote pullquote-right">
"Experience becomes an asset when it is structured, indexed, and retrievable."
</div>
<h4>Example Use Cases</h4>
<ul>
<li>A record that a promotion increased sales by 20 percent last Black Friday</li>
<li>A past fraud alert where a large transaction at midnight was confirmed as suspicious</li>
</ul>
<h3>3. Semantic Memory</h3>
<p>Semantic memory represents generalised knowledge. It is not tied to a single event, but rather to facts, rules, and domain understanding.</p>
<p>This is where agents transition from remembering to knowing.</p>
<ul>
<li>Encodes domain knowledge and concepts</li>
<li>Supports reasoning beyond immediate context</li>
<li>Often sourced from curated datasets or knowledge graphs</li>
<li>Enables abstraction across multiple experiences</li>
</ul>
<h4>Example Use Cases</h4>
<ul>
<li>The definition of a “high value customer” based on spend and frequency</li>
<li>The rule that international transactions carry higher fraud risk than local ones</li>
</ul>
<h2>How Memory Is Typically Stored</h2>
<p>The storage layer is where architectural decisions become tangible. Each memory type aligns to different data stores based on access patterns, latency requirements, and structure.</p>





























<table><thead><tr><th>Memory Type</th><th>Typical Storage Mechanism</th><th>Characteristics of Storage</th><th>Common Enterprise Technologies</th></tr></thead><tbody><tr><td>Short Term Memory</td><td>In memory cache</td><td>Ultra low latency, ephemeral, session scoped</td><td>Redis Cache, Memcached</td></tr><tr><td>Episodic (Long Term Memory)</td><td>Vector store + durable storage</td><td>Semantic retrieval combined with persistence across time</td><td>Pinecone, Weaviate, OpenSearch (vector), SQL Server, PostgreSQL</td></tr><tr><td>Semantic Memory</td><td>Document store / knowledge graph</td><td>Structured and unstructured knowledge, supports relationships and indexing</td><td>Elasticsearch, MongoDB, Neo4j</td></tr></tbody></table>
<p>There is a certain satisfaction in getting this layer right. When storage aligns with memory type, retrieval becomes predictable and performance follows naturally. When it does not, systems tend to behave like they have misplaced their keys, often at the least convenient moment.</p>
<h2>Memory Versus RAG</h2>
<p>Retrieval Augmented Generation, often referred to as RAG, plays an important role in modern AI systems. It provides access to external knowledge at the moment of inference. Yet it operates with a fundamentally different purpose compared to agent memory.</p>
<p>RAG focuses on enriching a single response. Memory focuses on shaping behaviour over time.</p>








































<table><thead><tr><th>Aspect</th><th>Agent Memory</th><th>RAG (Retrieval Augmented Generation)</th></tr></thead><tbody><tr><td>Purpose</td><td>Build continuity and learning across interactions</td><td>Enhance responses with external knowledge</td></tr><tr><td>Time Horizon</td><td>Persistent across sessions and workflows</td><td>Single query or request scope</td></tr><tr><td>Data Type</td><td>Experiences, knowledge, workflows, state</td><td>Documents, chunks, embeddings</td></tr><tr><td>Evolution</td><td>Continuously updated and refined</td><td>Typically static or periodically refreshed</td></tr><tr><td>Role in System</td><td>Core part of agent reasoning loop</td><td>Supporting retrieval layer for context injection</td></tr><tr><td>Behaviour Impact</td><td>Influences decisions, actions, and strategy</td><td>Improves factual grounding of outputs</td></tr></tbody></table>
<div class="quote">
"RAG helps an agent answer well. Memory helps an agent behave well."
</div>
<p>In practice, the two often work together. Episodic and semantic memory may be implemented using similar technologies as RAG, such as vector stores. The distinction lies in intent and orchestration.</p>
<p>RAG retrieves information. Memory integrates it into the agent’s ongoing state and decision process.</p>
<h2>Deciding What to Remember</h2>
<p>Designing memory is as much about restraint as it is about capability. The question is not simply what can be stored, but what should be stored.</p>
<p>Data within agent systems carries different lifecycles. Some information is highly transient, relevant only for a single interaction. Other data becomes stale as conditions change. A smaller subset holds long term value and can inform future decisions.</p>
<ul>
<li>Transient data should remain in short term memory and expire quickly</li>
<li>Stale data requires active management, including decay or revalidation</li>
<li>Expired or outdated knowledge can degrade decision quality if reused</li>
<li>Sensitive data introduces privacy and regulatory considerations</li>
</ul>
<div class="pullquote pullquote-right">
"An agent that remembers everything eventually trusts the wrong thing."
</div>
<p>Privacy becomes particularly important when memory persists across sessions. Personal or sensitive information must be handled with strict controls, including minimisation, anonymisation, and clear retention policies.</p>
<p>Equally, the reuse of stale or contextually incorrect data introduces subtle risks. Decisions may appear consistent while quietly drifting away from reality. This is where governance, lifecycle management, and observability play a crucial role in maintaining trust in the system.</p>
<h2>References</h2>
<p>Tulving, E. (1972). Episodic and Semantic Memory<br>
Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks<br>
OpenAI (2024–2025). Function Calling, Memory, and Tool Use in LLMs<br>
LangChain / LangGraph (2024–2025). Agent Memory and State Management Documentation</p>

<p>Hospitals are moving from AI as a clever demo toward AI as an operational teammate. That shift deserves stronger evidence. A randomised controlled trial can help a hospital test whether an agent genuinely improves flow, documentation, education, or discharge quality in the messy reality of day to day care, rather than only looking impressive in a pilot deck.<sup><a href="#user-content-fn-1" id="user-content-fnref-1" data-footnote-ref="" aria-describedby="footnote-label">1</a></sup><sup><a href="#user-content-fn-2" id="user-content-fnref-2" data-footnote-ref="" aria-describedby="footnote-label">2</a></sup></p>
<p>For operational agents, the most useful design is often a <strong>pragmatic cluster randomised trial</strong>, and in many hospitals the most practical version is a <strong>stepped wedge</strong> rollout. In that design, groups such as wards, teams, or clinics are randomised to the order in which they receive the new agent. Everyone starts in usual care, then clusters cross over to the agent in a random order over time. It is rigorous, fair, and well suited to hospitals where a gradual rollout is already the sensible path.<sup><a href="#user-content-fn-3" id="user-content-fnref-3" data-footnote-ref="" aria-describedby="footnote-label">3</a></sup><sup><a href="#user-content-fn-4" id="user-content-fnref-4" data-footnote-ref="" aria-describedby="footnote-label">4</a></sup></p>
<div class="quote">
“An AI agent in a hospital should earn trust the same way any new service does, by improving outcomes in the real world, with the humans still firmly in charge.”
</div>
<h2>Types of trials you can use</h2>
<p>Before designing the trial, it helps to understand the main types of randomised trials used in hospitals. Each one answers a slightly different question depending on how the AI agent fits into the workflow.</p>
<details><summary>Individual randomised controlled trial</summary><div class="accordion-content">
<p>This is the simplest form of a trial. Each patient is randomly assigned to either receive the AI supported care or the usual care.</p>
<p>You can think of it like flipping a coin for each patient. Some get the agent, some do not, and then you compare outcomes between the two groups.</p>
<p>This works best when the intervention is isolated to one patient and does not change how the wider team behaves.</p>
<ul>
<li>Example: A patient education agent explaining a procedure</li>
<li>Easy to analyse and very clean statistically</li>
<li>Risk of “spillover” if staff change behaviour for all patients</li>
</ul>
</div></details>
<details><summary>Cluster randomised trial</summary><div class="accordion-content">
<p>Instead of randomising individual patients, you randomise groups. These groups are called clusters and could be wards, teams, or clinics.</p>
<p>This is useful when the AI agent changes how a team works. If one nurse uses a discharge agent, it will likely influence how the whole ward operates.</p>
<p>So rather than mixing patients within the same ward, you assign entire wards to either use the agent or continue as usual.</p>
<ul>
<li>Example: One ward uses a discharge agent, another does not</li>
<li>Reduces contamination between groups</li>
<li>Slightly more complex to analyse because patients in the same cluster behave similarly</li>
</ul>
</div></details>
<details><summary>Stepped wedge trial</summary><div class="accordion-content">
<p>This is a special type of cluster trial where everyone eventually gets the AI agent, just at different times.</p>
<p>Imagine a rollout plan. At the start, no wards use the agent. Then, one ward starts using it. Later, another ward joins. Over time, all wards adopt it, but in a random order.</p>
<p>This allows you to compare:</p>
<ul>
<li>Wards before and after they receive the agent</li>
<li>Wards that have the agent versus those that do not yet have it at the same time</li>
</ul>
<p>It works well in hospitals because:</p>
<ul>
<li>
<p>It feels fair since everyone gets access</p>
</li>
<li>
<p>It matches how systems are usually rolled out</p>
</li>
<li>
<p>It captures real world operational conditions</p>
</li>
<li>
<p>Example: Bed allocation agent introduced ward by ward over several months</p>
</li>
</ul>
</div></details>
<details><summary>Pragmatic trial</summary><div class="accordion-content">
<p>A pragmatic trial focuses on testing the AI agent in real world conditions, rather than under perfect or controlled settings.</p>
<p>Instead of tightly controlling everything, you allow normal variation in staff, workload, and behaviour. This gives results that reflect how the agent will actually perform in practice.</p>
<ul>
<li>Example: A documentation agent used during normal clinics with real patients and time pressures</li>
<li>High relevance for decision makers</li>
<li>Slightly more noise in the data, but more realistic insights</li>
</ul>
</div></details>
<h2>Why randomise at all</h2>
<p>The case for randomisation is simple. Hospital operations change every week. Winter pressure arrives, staffing shifts, new consultants join, wards fill, and teams naturally get better at using any new tool over time. If a hospital simply compares before and after, it can end up crediting the agent for changes that were already on the way.</p>
<p>A randomised design creates a cleaner comparison. It helps separate the effect of the agent from the effect of timing, local enthusiasm, and workflow variation. For AI, that matters even more because guidance such as SPIRIT AI and CONSORT AI asks teams to be explicit about the model, the workflow, the human oversight, the input data, and how errors are handled.<sup><a href="#user-content-fn-1" id="user-content-fnref-1-2" data-footnote-ref="" aria-describedby="footnote-label">1</a></sup><sup><a href="#user-content-fn-2" id="user-content-fnref-2-2" data-footnote-ref="" aria-describedby="footnote-label">2</a></sup> Recent reviews also show that reporting quality in AI trials still has room to grow, which makes careful trial design even more valuable.<sup><a href="#user-content-fn-5" id="user-content-fnref-5" data-footnote-ref="" aria-describedby="footnote-label">5</a></sup></p>
<p>A good RCT also gives executives and clinical leaders answers they can act on:</p>
<ul>
<li>Did the agent improve the primary operational or clinical outcome?</li>
<li>Did it save staff time without shifting work somewhere else?</li>
<li>Did it behave safely across different wards and patient groups?</li>
<li>Did the benefit hold once the novelty wore off?</li>
</ul>
<h2>The best trial shape for hospital agents</h2>
<p>For a <strong>bed allocation agent</strong> or <strong>discharge agent</strong>, randomising individual patients is usually awkward because the intervention changes team behaviour and shared operational decisions. A ward, service line, or bed management team is often the better unit of randomisation. That reduces contamination and mirrors the way the work is actually done.<sup><a href="#user-content-fn-3" id="user-content-fnref-3-2" data-footnote-ref="" aria-describedby="footnote-label">3</a></sup></p>
<p>For a <strong>clinical documentation agent</strong>, there are two credible options. One is to randomise clinicians. The other is to randomise clinic sessions or care teams. A recent pragmatic randomised trial of ambient AI scribes assigned 238 physicians across 14 specialties to one of two scribe tools or usual care and measured time in note and burnout related outcomes, which is exactly the sort of practical evaluation hospitals need more of.<sup><a href="#user-content-fn-6" id="user-content-fnref-6" data-footnote-ref="" aria-describedby="footnote-label">6</a></sup></p>
<p>For a <strong>patient education agent</strong>, individual patient randomisation can work very well because the interaction is easier to isolate. There is already randomised work in this area, and broader reviews suggest strong potential, while also highlighting the importance of trust, privacy, and good behavioural design.<sup><a href="#user-content-fn-7" id="user-content-fnref-7" data-footnote-ref="" aria-describedby="footnote-label">7</a></sup><sup><a href="#user-content-fn-8" id="user-content-fnref-8" data-footnote-ref="" aria-describedby="footnote-label">8</a></sup></p>
<details><summary>Choosing the unit of randomisation</summary><div class="accordion-content">
<p>A useful rule is to randomise at the level where the workflow actually changes.</p>
<ul>
<li><strong>Patient level</strong> when the agent interacts directly with one patient and does not alter the wider team process</li>
<li><strong>Clinician level</strong> when the agent supports one doctor, nurse, or allied professional</li>
<li><strong>Team or ward level</strong> when the agent changes shared operations, prioritisation, or handoffs</li>
<li><strong>Site level</strong> when the intervention depends on local governance, infrastructure, or major process redesign</li>
</ul>
</div></details>
<h2>How to run the trial</h2>
<p>Start by defining the agent in operational terms. That means its exact task, where it appears in the workflow, what data it sees, what it recommends or drafts, who approves the action, and what happens when the output is incomplete. In hospitals, clarity is a safety feature disguised as project management.</p>
<p>Then choose one primary outcome only. You may admire twelve dashboards, but the statisticians will sleep better with one main question.</p>
<h3>1. Bed allocation agent</h3>
<p>A bed allocation agent might suggest the best next bed for an admitted patient based on specialty fit, infection control, acuity, cleaning status, and predicted discharges. The primary outcome could be <strong>time from admission decision to bed assignment</strong>. Secondary outcomes could include boarding time, number of bed moves, outlier placements, and staff override rate.</p>
<p>A sensible design would be a stepped wedge trial across wards or bed management shifts.</p>
<ul>
<li>Randomise wards to the order of activation</li>
<li>Keep human bed managers as final decision makers</li>
<li>Log every recommendation, override, and reason</li>
<li>Review equity by age, frailty, specialty, and protected characteristics where available</li>
</ul>
<h3>2. Discharge agent</h3>
<p>A discharge agent might assemble discharge tasks, identify blockers, draft summaries, and nudge teams when tests, pharmacy, transport, or community referrals are the remaining delay. The primary outcome could be <strong>time from medically fit decision to actual discharge</strong>. Secondary outcomes could include discharge before noon, seven day readmission, pharmacy turnaround, and patient experience.</p>
<p>This often suits cluster randomisation by ward or consultant team.</p>
<ul>
<li>Use the same discharge criteria in both arms</li>
<li>Measure whether work is genuinely completed earlier, not merely documented earlier</li>
<li>Track readmissions and post discharge contacts as balancing measures</li>
<li>Include manual safety review for complex cases</li>
</ul>
<h3>3. Clinical documentation agent</h3>
<p>A documentation agent may listen, transcribe, and draft notes or letters for clinician review. Here the primary outcome is often <strong>time in note</strong> or <strong>same day note closure</strong>. Secondary outcomes might include clinician workload, patient satisfaction, note quality, coding completeness, and error rate.</p>
<p>The recent ambient scribe trial offers a strong template: pragmatic deployment, routine workflow measurement, and survey based staff outcomes alongside safety monitoring.<sup><a href="#user-content-fn-6" id="user-content-fnref-6-2" data-footnote-ref="" aria-describedby="footnote-label">6</a></sup></p>
<ul>
<li>Randomise clinicians or clinic sessions</li>
<li>Use structured note quality review on a sample</li>
<li>Record edits, rejections, and clinically meaningful inaccuracies</li>
<li>Keep approval and sign off with the responsible clinician</li>
</ul>
<h3>4. Patient education agent</h3>
<p>A patient education agent may explain a procedure, medicine, prep instructions, or discharge plan in plain language and multiple languages. The primary outcome could be <strong>knowledge gain</strong> or <strong>adherence to the intended next step</strong>. Secondary outcomes could include anxiety, call backs, digital engagement, and satisfaction.</p>
<p>This can often be tested at patient level.</p>
<ul>
<li>Randomise eligible patients to usual education or usual education plus the agent</li>
<li>Measure comprehension after the visit and again later</li>
<li>Design for readability, culture, and accessibility</li>
<li>Escalate uncertain or clinical questions to a human</li>
</ul>
<div class="pullquote pullquote-right">
The elegant part of a hospital RCT is that it respects reality. The glamorous part is mostly the spreadsheet.
</div>
<h2>What must be stored and what should fade away</h2>
<p>AI agents work better with memory, yet a trial should still be disciplined about what the system keeps. Hospitals should decide in advance what information is useful, what becomes stale, and what should expire. A bed allocation agent may benefit from recent occupancy, cleaning state, or specialty constraints, while a patient education agent may need the current leaflet version, preferred language, and the specific procedure pathway for that episode of care.</p>
<p>The trial protocol should set clear retention and review rules.</p>
<ul>
<li>Store only the minimum data needed for the agent task and the trial analysis</li>
<li>Expire transient operational signals when they are no longer relevant</li>
<li>Version prompts, policies, and knowledge sources so results stay interpretable</li>
<li>Separate operational memory from patient record truth and from research logs</li>
</ul>
<p>This matters for privacy, governance, and scientific clarity. It also saves teams from the deeply modern problem of an agent confidently acting on information that was perfectly sensible last Thursday.</p>
<h2>Safety and fairness during the trial</h2>
<p>Every hospital AI trial should include balancing measures and active monitoring. A discharge agent that speeds discharge while increasing readmissions has taught you something useful, just perhaps not what the brochure promised. A documentation agent that saves six minutes but introduces occasional inaccuracies needs a careful view of net value. In one recent study of AI generated patient centred discharge instructions, potentially harmful safety issues were identified in 18 percent of outputs, including hallucinations and new medications, which is a strong reminder that polished language and safe content are very different things.<sup><a href="#user-content-fn-9" id="user-content-fnref-9" data-footnote-ref="" aria-describedby="footnote-label">9</a></sup></p>
<p>Plan governance from day one.</p>
<ul>
<li>Predefine stop rules for safety signals</li>
<li>Audit overrides because they reveal where the model and workflow misalign</li>
<li>Review subgroup performance, especially for language and complexity</li>
<li>Freeze important model versions during the main trial period where feasible</li>
</ul>
<h2>What success looks like</h2>
<p>A successful trial does more than show a p value. It gives the hospital a deployment decision. It tells leaders whether the agent improves the chosen outcome, where it works best, which teams trust it, how much human oversight it needs, and whether the gain is large enough to justify adoption at scale.</p>
<p>That is the real prize. An RCT for hospital AI is not about slowing innovation down. It is how a hospital turns experimentation into dependable capability, one ward, one team, and one measured decision at a time.</p>
<section data-footnotes="" class="footnotes"><h2 class="sr-only" id="footnote-label">Footnotes</h2>
<ol>
<li id="user-content-fn-1">
<p>Liu X, Cruz Rivera S, Moher D, et al. <em>CONSORT AI extension</em>. Nature Medicine. 2020. Reporting guideline for clinical trial reports for interventions involving artificial intelligence. <a href="#user-content-fnref-1" data-footnote-backref="" aria-label="Back to reference 1" class="data-footnote-backref">↩</a> <a href="#user-content-fnref-1-2" data-footnote-backref="" aria-label="Back to reference 1-2" class="data-footnote-backref">↩<sup>2</sup></a></p>
</li>
<li id="user-content-fn-2">
<p>Cruz Rivera S, Liu X, Chan A W, et al. <em>SPIRIT AI extension</em>. Nature Medicine. 2020. Protocol guidance for trials of interventions involving artificial intelligence. <a href="#user-content-fnref-2" data-footnote-backref="" aria-label="Back to reference 2" class="data-footnote-backref">↩</a> <a href="#user-content-fnref-2-2" data-footnote-backref="" aria-label="Back to reference 2-2" class="data-footnote-backref">↩<sup>2</sup></a></p>
</li>
<li id="user-content-fn-3">
<p>The Stepped Wedge Clinical Trial. <em>JAMA Guide to Statistics and Methods</em>. Overview of stepped wedge cluster randomised design and reasons for its use. <a href="#user-content-fnref-3" data-footnote-backref="" aria-label="Back to reference 3" class="data-footnote-backref">↩</a> <a href="#user-content-fnref-3-2" data-footnote-backref="" aria-label="Back to reference 3-2" class="data-footnote-backref">↩<sup>2</sup></a></p>
</li>
<li id="user-content-fn-4">
<p>Zhang S, Cui W, Ding S, et al. <em>A cluster randomized controlled trial of a nurse led artificial intelligence assisted prevention and management for delirium</em>. PLOS One. 2024. <a href="#user-content-fnref-4" data-footnote-backref="" aria-label="Back to reference 4" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-5">
<p>Martindale APL, et al. <em>Concordance of randomised controlled trials for artificial intelligence interventions with the CONSORT AI reporting guidelines</em>. Nature Communications. 2024. <a href="#user-content-fnref-5" data-footnote-backref="" aria-label="Back to reference 5" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-6">
<p>Lukac PJ, Turner W, Vangala S, et al. <em>Ambient AI Scribes in Clinical Practice</em>. NEJM AI. 2025. <a href="#user-content-fnref-6" data-footnote-backref="" aria-label="Back to reference 6" class="data-footnote-backref">↩</a> <a href="#user-content-fnref-6-2" data-footnote-backref="" aria-label="Back to reference 6-2" class="data-footnote-backref">↩<sup>2</sup></a></p>
</li>
<li id="user-content-fn-7">
<p>Su B, et al. <em>Chatbot for patient education for prenatal aneuploidy testing</em>. Patient Education and Counseling. 2025. <a href="#user-content-fnref-7" data-footnote-backref="" aria-label="Back to reference 7" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-8">
<p>Choi KEA, Fitzek S. <em>User and Provider Experiences With Health Education Chatbots</em>. JMIR Human Factors. 2025. <a href="#user-content-fnref-8" data-footnote-backref="" aria-label="Back to reference 8" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-9">
<p>Stanceski K, et al. <em>The quality and safety of using generative AI to produce patient centred discharge instructions</em>. npj Digital Medicine. 2024. <a href="#user-content-fnref-9" data-footnote-backref="" aria-label="Back to reference 9" class="data-footnote-backref">↩</a></p>
</li>
</ol>
</section>

<h2>For thirty years, software development has operated on a simple, unspoken hierarchy. Specifications were inputs. Code was the output. The document you approved in a strategy meeting became a rough guide — a wish list that developers interpreted, adapted, and ultimately rewrote in the language of implementation. By the time software shipped, the original intent had been filtered through dozens of technical decisions that no one above engineering had visibility into or control over. That hierarchy is changing. And the implications for how you lead, govern, and invest in technology are significant.</h2>
<hr>
<h2>The Gap That Has Always Existed</h2>
<p>Every technology leader knows the scenario. A detailed requirements document is agreed. Development begins. Months later, what ships doesn't quite match the original vision. Features are missing. Edge cases were interpreted differently. The roadmap shifts. The team embarks on another sprint cycle to correct course.</p>
<p>This isn't a failure of process or of people. It is a structural problem baked into how software has always been built.</p>
<p>The specification served the code. It was scaffolding — useful during planning, discarded once development began. Code became the source of truth. The original intent lived in documents that quickly fell out of sync with the system being built.</p>
<div class="quote">
"The gap between specification and implementation has plagued software development since its inception. We've tried to bridge it with better documentation, more detailed requirements, stricter processes. These approaches fail because they accept the gap as inevitable."
</div>
<p><strong>Spec-Driven Development (SDD)</strong> — a methodology formalised by GitHub's emerging spec-kit tooling — inverts this entirely. Code no longer drives specifications. Specifications drive code. The document you and your team agree on becomes the executable source of truth from which code is generated, tested, and maintained.</p>
<hr>
<h2>How SDD Differs From What Came Before</h2>
<p>To understand what SDD changes, it helps to place it against the methodologies executives are already familiar with.</p>
<h3>Waterfall</h3>
<p>Waterfall came closest to placing specifications at the centre. Requirements were documented in detail before a line of code was written. But Waterfall's fatal flaw was that specifications and code were still entirely separate artefacts. Once requirements were handed to engineering, the spec became static. It did not evolve with the product. It did not generate code. It simply described what was intended, and engineers translated that intent manually — a translation process riddled with interpretation, assumption, and drift.</p>
<p>When reality met the specification, it was always the specification that lost.</p>
<h3>Agile</h3>
<p>Agile was a direct reaction to the failures of Waterfall — and it made a reasonable bet. If specifications were always going to fall behind reality, stop trying to write them upfront. Embrace change. Iterate. Let working software be the measure of progress.</p>
<p>The result was a methodology that elevated developer velocity and shortened feedback cycles. But it came at a cost that is only now becoming fully visible: <strong>specifications became secondary artefacts</strong>. Product backlogs, user stories, and acceptance criteria served the sprint. They described intent, but code was still the truth. When requirements changed, teams manually propagated those changes through design documents, APIs, tests, and implementations. Each pivot accumulated hidden technical debt.</p>
<div class="grey-box">
<p><strong>The Agile Trade-off</strong></p>
<p>Agile solved the delivery problem but preserved the intent-to-implementation gap. It made that gap faster to navigate, not smaller. The backlog remained a proxy for business intent, not a driver of it.</p>
</div>
<h3>SAFe and Scaled Frameworks</h3>
<p>Enterprise frameworks like SAFe attempted to bring strategic alignment back into Agile at scale. Program Increments, Architectural Runways, and Portfolio Kanban boards tried to reconnect business goals to engineering work. But they added process overhead without resolving the core issue: the translation between business intent and working code still happened manually, inside individual teams, with all the interpretation and drift that implies.</p>
<hr>
<h2>What SDD Actually Changes</h2>
<p>SDD is made possible by a single recent shift: AI systems can now reliably translate precise natural-language specifications into working code. That capability breaks the assumption that manual translation was inevitable.</p>
<p>In an SDD workflow, specifications are not inputs to development. They are the development. A Product Requirements Document is not a guide for engineers — it is the artefact from which implementation plans are generated, and from which code is produced. When you change a requirement in the PRD, affected technical decisions update. When you modify a user story, corresponding API endpoints regenerate. The document and the system stay in sync because the document produces the system.</p>
<div class="quote">
"When specifications drive implementation, pivots become systematic regenerations rather than manual rewrites."
</div>
<h3>Three Stages of the SDD Workflow</h3>
<p><strong>1. Specification</strong> — A product idea is developed iteratively with AI assistance into a precise, structured PRD. AI asks clarifying questions, identifies edge cases, and flags ambiguities with explicit markers rather than guessing. What would traditionally require days of meetings and documentation can happen in hours.</p>
<p><strong>2. Implementation Planning</strong> — From the PRD, AI generates a detailed implementation plan. Every technology choice carries documented rationale. Every architectural decision traces back to a specific requirement. Organisational constraints — security policies, database standards, authentication requirements — are applied automatically.</p>
<p><strong>3. Code Generation and Testing</strong> — Code is generated from the implementation plan. Crucially, test scenarios are part of the specification itself, not an afterthought. Tests define expected behaviour before code is written, ensuring that what gets built matches what was specified.</p>
<hr>
<h2>The Business Case for Executives</h2>
<details>
<summary>Faster, More Predictable Pivots</summary>
<div class="accordion-content">
<p>In traditional development, a significant requirement change triggers a cascade of manual updates across documentation, design, and code. Each manual step introduces lag and the risk of inconsistency. SDD transforms this into a systematic process: change the specification, regenerate the affected plan, regenerate the affected code. The gap between decision and implementation compresses from weeks to hours.</p>
<p>For businesses operating in competitive, fast-moving markets, this is a structural advantage.</p>
</div>
</details>
<details>
<summary>Intent Preserved as an Asset</summary>
<div class="accordion-content">
<p>One of the least discussed costs of traditional development is the erosion of institutional intent. The original rationale behind architectural decisions disappears into git commit messages and the memory of engineers who may no longer be on the team. Specifications and code diverge. The business no longer has a readable, accurate account of why the system is the way it is.</p>
<p>SDD keeps intent visible and current. The specification is always accurate because it generates the code. When new team members join, auditors arrive, or regulators ask questions, the answer to "why does the system work this way?" lives in a document that is both readable and up to date.</p>
</div>
</details>
<details>
<summary>Reduced Technical Debt at the Source</summary>
<div class="accordion-content">
<p>Technical debt is largely a product of manual translation. Engineers make reasonable decisions in the moment that diverge from original intent. Those decisions accumulate. Over time, the cost of change rises. SDD addresses this at the source. Because code is generated from specifications, refactoring means updating the specification — not untangling years of accumulated interpretation.</p>
</div>
</details>
<details>
<summary>A New Basis for Governance</summary>
<div class="accordion-content">
<p>For CIOs and CTOs, SDD offers something previous methodologies could not: a single, auditable source of truth that bridges business intent and technical implementation. Compliance requirements, security constraints, and architectural principles can be encoded directly into the specification layer, ensuring that every generated system automatically inherits them.</p>
</div>
</details>
<div class="grey-box">
<p><strong>Constitutional Governance</strong></p>
<p>SDD introduces the concept of a development constitution — a set of immutable architectural principles that govern every specification and every generated implementation. Simplicity constraints, test-first mandates, integration requirements, and modularity rules are enforced at the specification level, not left to individual engineering discretion. For enterprises managing complex technology estates, this is governance infrastructure.</p>
</div>
<hr>
<h2>What Changes for Your Organisation</h2>
<p>SDD does not eliminate the need for engineering expertise. It redirects it. The shift is from mechanical translation — turning specifications into code by hand — toward specification quality. The most valuable skill in an SDD organisation is the ability to write precise, unambiguous, complete specifications that AI can execute reliably.</p>
<p>This changes the profile of what strong engineering leadership looks like. It elevates the role of product thinking, systems design, and clear expression of intent. It reduces the premium on the ability to hold large, complex codebases in memory.</p>
<hr>
<h2>Key Artefacts in a Spec-Driven Workflow</h2>
<p>SDD produces a set of structured Markdown documents that together form the complete, auditable record of a system's intent and design. Each artefact has a distinct role and is either authored by a human or generated by AI from an upstream document.</p>





















































<table><thead><tr><th>Artefact</th><th>Filename Convention</th><th>Purpose</th><th>Authored By</th></tr></thead><tbody><tr><td>Product Requirements Document</td><td><code>PRD.md</code></td><td>Captures business intent, user stories, scope, and constraints. The root of the specification chain.</td><td>Human (AI-assisted)</td></tr><tr><td>Implementation Plan</td><td><code>PLAN.md</code></td><td>Maps every PRD requirement to specific technical decisions, components, and architecture choices with documented rationale.</td><td>AI (from PRD)</td></tr><tr><td>Development Constitution</td><td><code>CONSTITUTION.md</code></td><td>Defines immutable architectural principles — simplicity rules, security constraints, integration mandates — applied to every downstream artefact.</td><td>Human</td></tr><tr><td>API Specification</td><td><code>API.md</code></td><td>Describes all interfaces, endpoints, request/response shapes, and error contracts derived from the PRD.</td><td>AI (from PLAN)</td></tr><tr><td>Test Scenarios</td><td><code>TESTS.md</code></td><td>Defines expected behaviour and acceptance criteria before code is written. Used to verify generated implementations.</td><td>AI (from PRD)</td></tr><tr><td>Architecture Decision Records</td><td><code>decisions/ADR-NNN.md</code></td><td>Captures individual architectural choices, alternatives considered, and the reasons for decisions. Generated inline during planning.</td><td>AI (from PLAN)</td></tr><tr><td>Changelog</td><td><code>CHANGELOG.md</code></td><td>Tracks specification changes and their downstream effects on plan and code, maintaining traceability over time.</td><td>AI (from diffs)</td></tr></tbody></table>
<p>The critical property of this artefact chain is directionality: changes flow downward from the PRD. Modifying a requirement regenerates the PLAN, which in turn regenerates affected API, test, and implementation artefacts. No manual propagation. No silent drift.</p>
<hr>
<h2>The Honest Caveats</h2>
<p>SDD is not a silver bullet, and it is not fully mature. Practising it today requires assembling tools, maintaining discipline, and accepting that the methodology is still evolving. AI code generation is reliable but not infallible. Specifications still require human judgement to write well. The feedback loop between production systems and specification updates is a discipline that teams need to build.</p>
<p>What SDD represents is not a solved problem but a directional shift. The organisations that learn to write excellent specifications now — precise, complete, unambiguous — will be the ones best positioned as AI generation becomes more capable and the tooling matures.</p>
<div class="quote">
"SDD provides the methodology for achieving alignment between intent and implementation through executable specifications that generate code rather than merely guiding it."
</div>
<hr>
<h2>References and Further Reading</h2>
<ul>
<li><strong>GitHub spec-kit — Spec-Driven Development Guide</strong> — The reference implementation of SDD tooling and workflow conventions. <a href="https://github.com/github/spec-kit/blob/main/spec-driven.md">github.com/github/spec-kit</a></li>
<li><strong>Agile Manifesto</strong> — <a href="https://agilemanifesto.org">agilemanifesto.org</a>. The original principles behind iterative development — and the implicit trade-offs that SDD now addresses.</li>
<li><strong>Scaled Agile Framework (SAFe)</strong> — <a href="https://scaledagileframework.com">scaledagileframework.com</a>. Enterprise Agile framework context for understanding where SDD sits relative to existing delivery structures.</li>
</ul>

<p>In the current wave of enterprise AI adoption, attention often gravitates toward model capability. Larger context windows, stronger reasoning, and improved tool use dominate the narrative. Yet beneath this visible progress, a quieter structural shift is unfolding. It is one that will define whether agentic systems scale safely inside healthcare organisations or introduce new forms of operational risk.</p>
<p>That shift is the emergence of <strong>agent identity as a first class architectural concern</strong>.</p>
<p>For CIOs and CTOs, this is a decision that directly impacts governance, auditability, and clinical safety.</p>
<hr>
<h2>From Intelligent Systems to Collaborative Agents</h2>
<p>Agentic AI introduces a change in how work is executed across systems. Instead of a single system completing a task end to end, we now see distributed execution across multiple specialised agents.</p>
<p>A clinical scenario illustrates this clearly.</p>
<p>A request is triggered to optimise patient discharge.</p>
<p>One agent retrieves bed state and occupancy data.<br>
Another evaluates discharge readiness using clinical indicators.<br>
A third generates recommended actions.<br>
A fourth validates compliance and policy constraints before execution.</p>
<p>What appears as a single workflow is, in reality, a <strong>network of collaborating agents</strong>.</p>
<div class="quote">
“Once agents begin delegating work, the system stops being a workflow and starts becoming an ecosystem.”
</div>
<p>This is where traditional orchestration models begin to strain. Static pipelines assume predefined steps. Agentic systems introduce dynamic delegation, where decisions are made at runtime.</p>
<p>And with that, a fundamental question emerges.</p>
<p><strong>Who is performing each action?</strong></p>
<hr>
<h2>Identity as the Anchor of Trust</h2>
<p>In healthcare, identity is already foundational. Every clinician, system, and service operates within a defined identity boundary. Access to patient data, prescribing rights, and clinical decision support all depend on it.</p>
<p>Agentic systems require the same discipline.</p>
<p>An agent cannot operate as an anonymous executor or inherit a generic service account. It must possess:</p>
<ul>
<li>A distinct identity</li>
<li>Explicitly scoped permissions</li>
<li>Traceable accountability across every action</li>
</ul>
<p>Without this, the system loses its ability to answer critical questions:</p>
<ul>
<li>Which agent accessed patient data?</li>
<li>Under what authority was a decision made?</li>
<li>How did a recommendation propagate across systems?</li>
</ul>
<hr>
<h2>The Healthcare Context</h2>
<p>Healthcare environments amplify the importance of identity due to regulatory and operational constraints.</p>
<p>Patient data is governed under strict frameworks such as:</p>
<ul>
<li>GDPR (UK and EU)</li>
<li>NHS Data Security and Protection Toolkit</li>
<li>ISO 27001 and clinical safety standards (DCB0129 / DCB0160)</li>
</ul>
<p>Agentic systems operating without clear identity boundaries risk breaching all three simultaneously.</p>
<p>Consider a multi agent pathway:</p>
<p>An admission optimisation agent delegates bed allocation to another agent. That agent accesses EHR data, evaluates capacity, and triggers downstream workflows.</p>
<p>Without identity propagation:</p>
<ul>
<li>Data access becomes untraceable</li>
<li>Clinical accountability becomes ambiguous</li>
<li>Audit trails fragment across systems</li>
</ul>
<p>With identity:</p>
<ul>
<li>Every interaction is attributable</li>
<li>Permissions remain enforced at each step</li>
<li>Audit chains remain intact end to end</li>
</ul>
<p>This is the difference between controlled automation and uncontrolled autonomy.</p>
<hr>
<h2>Identity Propagation Across Agent Chains</h2>
<p>The complexity increases when agents delegate to other agents. Identity must not only exist, it must <strong>propagate deterministically</strong> across the execution chain.</p>
<p>Each interaction must carry:</p>
<ul>
<li>The originating agent identity</li>
<li>Delegation context</li>
<li>Permission boundaries</li>
<li>Execution constraints</li>
</ul>
<p>This creates a verifiable chain of custody for every action.</p>
<p>Below is a simplified view of identity propagation in a healthcare agent workflow.</p>
<svg width="100%" height="260" viewBox="0 0 800 260" xmlns="http://www.w3.org/2000/svg" style="border: 1px solid white;">
  <rect x="20" y="80" width="160" height="60" fill="#b5151b"></rect>
  <text x="30" y="115" fill="white" font-size="12">Discharge Agent</text>
  <rect x="220" y="40" width="160" height="60" fill="white"></rect>
  <text x="230" y="75" fill="#111111" font-size="12">Bed Allocation Agent</text>
  <rect x="420" y="80" width="160" height="60" fill="#cccccc"></rect>
  <text x="430" y="115" fill="#111111" font-size="12">EHR System</text>
  <rect x="620" y="40" width="160" height="60" fill="#b5151b"></rect>
  <text x="630" y="75" fill="white" font-size="12">Compliance Agent</text>
  <line x1="180" y1="110" x2="220" y2="70" stroke="#ccc" stroke-width="2"></line>
  <line x1="380" y1="70" x2="420" y2="110" stroke="#ccc" stroke-width="2"></line>
  <line x1="580" y1="110" x2="620" y2="70" stroke="#ccc" stroke-width="2"></line>
  <rect x="0" y="230" width="800" height="30" fill="white"></rect>
  <text x="400" y="250" font-size="12" fill="black" text-anchor="middle" font-weight="bold">Identity + Delegation Context</text>
</svg>
<h2>From Orchestration to an Agent Capability Ecosystem</h2>
<p>Once identity is established, a second capability becomes possible.</p>
<p><strong>Discovery.</strong></p>
<p>Agents can begin to expose their capabilities as services:</p>
<ul>
<li>What tasks they perform</li>
<li>What systems they access</li>
<li>Under which permissions they operate</li>
</ul>
<p>This creates a structured environment where agents can discover and invoke each other dynamically.</p>
<details>
<summary>Agent Capability Model</summary>
<div class="accordion-content">
<p>An agent exposes:</p>
<ul>
<li>Capability description</li>
<li>Required inputs and outputs</li>
<li>Permission scope</li>
<li>Policy constraints</li>
<li>Cost and execution limits</li>
</ul>
<p>Other agents evaluate:</p>
<ul>
<li>Suitability for task</li>
<li>Trust level based on identity</li>
<li>Compliance with policy</li>
</ul>
<p>Delegation becomes a controlled decision rather than a hard coded instruction.</p>
</div>
</details>
<p>This is the beginning of what many describe as an <strong>agent marketplace</strong>.</p>
<p>Not commercial in nature, but an architectural philosophy.</p>
<hr>
<h2>The Emergence of an Agent to Agent Economy</h2>
<p>At scale, these interactions evolve into something more dynamic.</p>
<p>Work begins to flow across a network of specialised agents:</p>
<ul>
<li>Tasks are decomposed</li>
<li>Capabilities are discovered</li>
<li>Execution is delegated</li>
<li>Results are recomposed</li>
</ul>
<p>The system becomes an <strong>adaptive execution fabric</strong>.</p>
<div class="pullquote pullquote-right">
“Agents do not just execute workflows. They select collaborators.”
</div>
<p>In healthcare, this could manifest as:</p>
<ul>
<li>Cross hospital resource optimisation</li>
<li>Regional bed coordination agents</li>
<li>Diagnostic support agents collaborating with imaging systems</li>
<li>Operational agents balancing demand across care pathways</li>
</ul>
<p>This introduces measurable efficiency gains.</p>
<p>Industry studies indicate that AI driven automation can improve operational efficiency by <strong>20–40% in healthcare workflows</strong>, particularly in scheduling, bed management, and administrative processes<sup><a href="#user-content-fn-1" id="user-content-fnref-1" data-footnote-ref="" aria-describedby="footnote-label">1</a></sup>.</p>
<hr>
<h2>Governance</h2>
<p>An agent ecosystem without governance becomes unstable very quickly.</p>
<p>Key risks include:</p>
<ul>
<li>Uncontrolled delegation loops</li>
<li>Permission escalation across agents</li>
<li>Loss of auditability</li>
<li>Unbounded cost from recursive execution</li>
</ul>
<p>To manage this, organisations must enforce:</p>
<ul>
<li>Identity based access control for agents</li>
<li>Policy driven delegation constraints</li>
<li>Full audit trails across agent interactions</li>
<li>Cost and execution limits per agent</li>
</ul>
<p>Or, put more bluntly, agents require the same discipline as humans. Possibly more.</p>
<p>(And unlike humans, they do not complain about password rotation, which is a small but welcome mercy!)</p>
<hr>
<h2>Architectural Priorities for CIOs and CTOs</h2>
<p>The introduction of agent identity is a foundational layer that must be designed deliberately.</p>
<p>Key priorities include:</p>
<ul>
<li>
<p><strong>Agent Identity Framework</strong><br>
Define how agents are issued identities, authenticated, and governed</p>
</li>
<li>
<p><strong>Permission Boundaries</strong><br>
Ensure least privilege access across all systems and tools</p>
</li>
<li>
<p><strong>Identity Propagation Standards</strong><br>
Maintain traceability across multi agent workflows</p>
</li>
<li>
<p><strong>Audit and Observability</strong><br>
Capture full execution chains for compliance and debugging</p>
</li>
<li>
<p><strong>Policy Enforcement Layer</strong><br>
Apply clinical, operational, and financial constraints</p>
</li>
</ul>
<p>These are prerequisites for safe deployment in healthcare environments.</p>
<hr>
<h2>A Shift That Happens Gradually</h2>
<p>This transformation will not arrive through a single platform or model release. It will emerge incrementally as organisations mature their agentic architectures.</p>
<p>At first, agents assist with isolated tasks.<br>
Then they begin coordinating workflows.<br>
Then they begin collaborating with each other.</p>
<p>And at that point, identity moves from a background concern to a <strong>central control mechanism</strong>.</p>
<p>Because in an ecosystem where agents collaborate, delegate, and execute autonomously, the ability to answer <strong>who did what, and under what authority</strong> becomes the defining requirement.</p>
<p>Identity is what allows intelligence to scale without losing accountability.</p>
<section data-footnotes="" class="footnotes"><h2 class="sr-only" id="footnote-label">Footnotes</h2>
<ol>
<li id="user-content-fn-1">
<p>McKinsey &#x26; Company, <em>The future of healthcare operations with AI</em>, 2023. <a href="#user-content-fnref-1" data-footnote-backref="" aria-label="Back to reference 1" class="data-footnote-backref">↩</a></p>
</li>
</ol>
</section>

<p>There has been a quiet shift in how we build with large language models. Not long ago, most systems began with a prompt, perhaps a few tools, and a chain stitched together to produce something useful. It worked, often surprisingly well, and for many teams that was enough.</p>
<p>Then systems started to grow. What began as a simple interaction turned into multiple steps, dependencies, retries, and edge cases. At that point, the idea of a chain began to feel slightly optimistic, in the same way a whiteboard sketch sometimes survives longer than it should.</p>
<p>This is where the conversation moves from LangChain to LangGraph, and now to Google’s Agent Development Kit. Each represents a stage in the same evolution.</p>
<hr>
<h2>From Prompts to Systems</h2>
<p>LangChain made it possible to connect models to the world. It gave developers a way to integrate tools, retrieval, and reasoning without building everything from scratch. It was, and still is, the fastest way to get something working.</p>
<p>But as systems matured, a pattern emerged. Developers began to reimplement the same ideas repeatedly:</p>
<ul>
<li>State management</li>
<li>Multi step workflows</li>
<li>Conditional execution</li>
<li>Error handling and retries</li>
</ul>
<p>LangGraph appeared as a response to this, bringing structure to orchestration and making execution visible.</p>
<div class="quote">
“LangChain helps you start. LangGraph helps you control. ADK expects you to design.”
</div>
<p>Google ADK takes this further by assuming that agents are not experiments but systems that must behave predictably.</p>
<hr>
<h2>Understanding the Stack</h2>
<p>Rather than competing frameworks, these tools form a layered stack that reflects increasing maturity.</p>

























<table><thead><tr><th>Layer</th><th>Framework</th><th>Purpose</th></tr></thead><tbody><tr><td>Components</td><td>LangChain</td><td>Tools, LLMs, retrieval, integrations</td></tr><tr><td>Orchestration</td><td>LangGraph</td><td>State, control flow, execution graphs</td></tr><tr><td>Agent Runtime</td><td>Google ADK</td><td>Structured agents, contracts, workflows</td></tr></tbody></table>
<p>Each layer builds on the one below it, introducing more control and more responsibility.</p>
<hr>
<h2>LangChain: Where Ideas Take Shape</h2>
<p>LangChain remains widely used because it optimises for speed and flexibility. It allows developers to quickly connect LLMs to tools and data sources, creating useful applications with minimal overhead.</p>
<p>It offers:</p>
<ul>
<li>Rapid prototyping of LLM applications</li>
<li>Extensive integrations</li>
<li>Flexible agent patterns</li>
<li>Minimal upfront structure</li>
</ul>
<p>This flexibility is often exactly what is needed in early stages, when the goal is discovery rather than control.</p>
<hr>
<h2>LangGraph: Where Control Emerges</h2>
<p>LangGraph introduces explicit orchestration. Instead of relying on the model to decide everything, developers define a graph where nodes represent steps and edges define transitions.</p>
<p>This approach brings clarity:</p>
<ul>
<li>Execution paths are defined rather than inferred</li>
<li>State is explicit and inspectable</li>
<li>Complex workflows become manageable</li>
<li>Debugging becomes significantly easier</li>
</ul>
<div class="pullquote pullquote-right">
“At some point, every chain becomes a graph. The only question is whether you admit it.”
</div>
<p>LangGraph does not remove flexibility. It channels it into a form that can be reasoned about.</p>
<hr>
<h2>Google ADK: Where Structure Becomes Non Negotiable</h2>
<p>Google ADK builds on similar principles but takes a more opinionated approach. It treats agents as structured components within a broader system, with clear contracts and defined behaviour.</p>
<p>Key characteristics include:</p>
<ul>
<li>Strongly typed inputs and outputs</li>
<li>Explicit agent definitions</li>
<li>Structured workflows</li>
<li>A defined runtime model</li>
</ul>
<p>Where LangGraph provides a graph engine, ADK provides a framework for building consistent, production ready agent systems.</p>
<hr>
<h2>The Architectural Differences</h2>
<p>The distinctions between the three approaches become clearer when viewed together.</p>















































<table><thead><tr><th>Dimension</th><th>LangChain</th><th>LangGraph</th><th>Google ADK</th></tr></thead><tbody><tr><td>Primary Focus</td><td>Integration</td><td>Orchestration</td><td>Agent architecture</td></tr><tr><td>Abstraction Level</td><td>High level components</td><td>Low level control</td><td>Structured system design</td></tr><tr><td>Execution Model</td><td>Prompt driven</td><td>Graph driven</td><td>Workflow driven</td></tr><tr><td>State Handling</td><td>Implicit</td><td>Explicit shared state</td><td>Structured inputs and outputs</td></tr><tr><td>Typing</td><td>Loose</td><td>Optional</td><td>Strong emphasis</td></tr><tr><td>Production Readiness</td><td>Requires layering</td><td>Requires design</td><td>Built with production in mind</td></tr></tbody></table>
<hr>
<h2>Where Each Fits</h2>
<details>
<summary>LangChain: Early Exploration</summary>
<div class="accordion-content">
<p>LangChain is well suited to early stage development where speed and flexibility matter. It allows teams to test ideas, integrate tools quickly, and understand what works before introducing structure.</p>
</div>
</details>
<details>
<summary>LangGraph: Workflow Orchestration</summary>
<div class="accordion-content">
<p>LangGraph becomes valuable when systems require control over execution. It supports complex workflows, branching logic, and stateful interactions that go beyond simple chains.</p>
</div>
</details>
<details>
<summary>Google ADK: Production Systems</summary>
<div class="accordion-content">
<p>ADK is designed for production environments where reliability, governance, and maintainability are required. It enforces structure and aligns with established engineering practices.</p>
</div>
</details>
<hr>
<h2>The Direction of Travel</h2>
<p>What is emerging is a layered engineering model for agents:</p>
<ul>
<li>LangChain lowers the barrier to entry</li>
<li>LangGraph introduces control and visibility</li>
<li>ADK enforces structure and consistency</li>
</ul>
<p>Each stage reflects increasing maturity in how systems are designed and operated.</p>
<p>There is also a moment many teams recognise. A system that began as a simple chain becomes increasingly complex, with hidden state and unpredictable behaviour. Introducing structure tends to resolve this quickly, even if it requires revisiting earlier decisions.</p>
<hr>
<h2>References</h2>
<ol>
<li>Google Agent Development Kit Documentation — <a href="https://google.github.io/adk-docs/">https://google.github.io/adk-docs/</a></li>
<li>LangChain Documentation — <a href="https://python.langchain.com">https://python.langchain.com</a></li>
<li>LangGraph Documentation — <a href="https://langchain-ai.github.io/langgraph/">https://langchain-ai.github.io/langgraph/</a></li>
</ol>

<p>This morning it was a flickering neon strip in the garage. There was a faint buzz, an occasional strobe effect, and just enough inconsistency to make it irritating rather than urgent. The sort of problem you feel you should be able to solve.</p>
<p>So I did what most of us now do. I took a photo, asked AI, and within seconds had a confident answer. Likely ballast failure. It even explained what a ballast was, how it worked, and what needed replacing.</p>
<p>Which is when I realised something slightly uncomfortable.</p>
<p>I had no idea what a ballast actually was. I had always dealt with the old setup. When a tube flickered, you replaced the starter. Simple, visible, and reassuringly mechanical. This was something else entirely.</p>
<hr>
<h2>The Illusion of Knowing</h2>
<p>Armed with a diagnosis, I opened the fitting expecting a quick fix. Instead, I found a sealed unit, labelled and wired in a way that immediately felt less familiar. AI had given me the answer, but it had not given me the instinct for what I was looking at.</p>
<div class="quote">
“AI can tell you what is wrong. It cannot tell you what it feels like to fix it.”
</div>
<p>On the screen, everything made sense. The explanation was clear, the steps were logical, and the outcome felt achievable. In reality, it was different. The wires, the layout, the small variations that never show up in diagrams all introduced hesitation.</p>
<hr>
<h2>The Gap Between Knowing and Doing</h2>
<p>This is where the gap becomes obvious. There is a difference between understanding a process and being able to carry it out with confidence. One is cognitive. The other is physical, situational, and built over time.</p>
<p>A trained electrician would open that fitting and immediately recognise the setup. They would not need to interpret each step. They would not second guess the wiring. They would not pause to check whether something looked right. They would simply do it.</p>
<p>That confidence is not explained. It is accumulated.</p>
<hr>
<h2>AI as a Diagnostic Layer</h2>
<p>AI is exceptionally good at compressing knowledge. It can identify faults, explain systems, and guide decisions in a way that feels almost complete. It removes the uncertainty of not knowing what the problem is.</p>
<p>But it stops just short of execution.</p>
<p>It does not give you the muscle memory of handling components. It does not give you the pattern recognition that comes from seeing the same issue dozens of times. It does not give you the quiet certainty that something is correct without needing to double check.</p>
<div class="pullquote pullquote-right">
“We are becoming fluent in explanation, while slowly losing fluency in execution.”
</div>
<hr>
<h2>A Wider Pattern</h2>
<p>This is not about lighting. It is about something broader.</p>
<p>We are optimising for people who can:</p>
<ul>
<li>Diagnose problems quickly</li>
<li>Ask precise questions</li>
<li>Navigate complex information</li>
</ul>
<p>At the same time, there is less emphasis on the people who can:</p>
<ul>
<li>Apply that knowledge in the real world</li>
<li>Work through imperfect conditions</li>
<li>Deliver outcomes with their hands</li>
</ul>
<p>The balance is shifting, and it is subtle enough that we rarely notice it.</p>
<hr>
<h2>A Quiet Realisation</h2>
<p>Standing there with the fitting open, I had everything I needed in theory. The diagnosis was correct. The explanation was clear. The steps were available.</p>
<p>What I lacked was the confidence to execute.</p>
<p>Not because the task was especially complex, but because I had never done it before. There is a difference between following instructions and knowing what you are doing. That difference becomes very obvious the moment something matters.</p>
<hr>
<h2>The Value of Specialists</h2>
<p>The person who replaces that ballast will not need to ask AI. They will not need to interpret the situation step by step. They will rely on something far more valuable.</p>
<p>Experience.</p>
<p>Years of seeing similar problems, making small adjustments, and developing an instinct for what works. That kind of capability does not scale through explanation. It scales through practice.</p>
<hr>
<h2>A Small Problem, A Larger Signal</h2>
<p>A flickering light is a minor inconvenience. It is also a useful signal.</p>
<p>AI is making us incredibly effective at understanding problems. It is raising the baseline of what we can diagnose and discuss. That is a genuine step forward.</p>
<p>At the same time, the ability to act on that understanding remains rooted in human skill, repetition, and judgement.</p>
<div class="quote">
“In a world of perfect answers, the rare skill is still knowing how to act on them.”
</div>
<hr>
<h2>The Practical Detail</h2>
<p>For reference, the fitting used an electronic ballast rather than the older starter based setup I was familiar with. That small shift in technology was enough to move the task from familiar to uncertain.</p>
<p>And that is the point.</p>
<p>The knowledge was available instantly. The execution was not.</p>
<p>Somewhere between those two sits a category of skill that is becoming quietly more valuable.</p>
<p>Not because it is new.</p>
<p>Because fewer people are practising it.</p>

<h2>At the <strong>London Tech Show 2026 at ExCeL</strong>, the stage lights turned to <strong>Baroness Martha Lane Fox</strong>, one of the UK’s most influential technology leaders and a co founder of lastminute.com. Martha has lived through the entire arc of the modern internet, from the early ecommerce experiments of the late 1990s to today’s rapidly evolving AI landscape.</h2>
<p>Her talk offered something rare. Perspective.</p>
<p>Most discussions about artificial intelligence start with the future. Martha began somewhere more useful. The past.</p>
<hr>
<h2>When the internet was still a strange idea</h2>
<p>In 1998 Martha was helping build <strong>lastminute.com</strong>, one of the earliest online travel businesses. The technology itself was not the hardest part of the job.</p>
<p>The real challenge was persuading people that buying something online was even possible.</p>
<div class="quote">
The challenge wasn't building the business. The challenge was persuading people they could safely buy something on the internet.
</div>
<p>Customers genuinely worried that entering their credit card details into a website might cause the money to vanish into the ether.</p>
<p>Today that concern feels almost charming. Online payments are routine. Digital services underpin entire industries. Yet at the time it felt radical.</p>
<p>The late nineties carried a sense of momentum. The UK technology scene was buzzing. New companies were emerging. The internet felt like an open frontier.</p>
<p>And, like all frontiers, it came with a few unexpected stories.</p>
<hr>
<h2>A lesson in unintended consequences</h2>
<p>One early feature on lastminute.com copied something Amazon had pioneered. One click purchasing.</p>
<p>The feature stored customer information so future bookings could happen instantly. Convenient and efficient.</p>
<p>Until one customer phoned customer services furious.</p>
<p>He had booked a trip using details saved from a previous holiday. Unfortunately the earlier trip had not been with his wife.</p>
<div class="pullquote pullquote-right">
He blamed me for the breakup of his marriage. That felt like quite a stretch.
</div>
<p>Technology moves quickly. Human life sometimes catches up a little more slowly.</p>
<hr>
<h2>When optimism met reality</h2>
<p>The dot com era eventually hit turbulence. Markets shifted. Companies disappeared. Businesses that had strong foundations survived while others quietly vanished.</p>
<p>Around the same time Martha experienced a personal turning point. A serious illness kept her in hospital for an extended period and offered a very different view of technology’s role in society.</p>
<p>Lying in hospital, surrounded by paper records and fragmented information systems, she saw how uneven the digital revolution had been.</p>
<p>The world she had come from was built on databases, services and real time information. The systems around her looked very different.</p>
<p>Technology had transformed commerce.</p>
<p>Many public services still had a long journey ahead.</p>
<hr>
<h2>The unfinished digital transformation</h2>
<p>Martha later helped shape the <strong>UK Government Digital Service</strong>, an effort to modernise how public services are delivered.</p>
<p>That work revealed something important.</p>
<p>Digital progress had not reached everyone equally.</p>
<p>During the COVID years this became painfully visible. Families shared a single phone so children could attend online lessons. Some households chose between buying mobile data or buying food.</p>
<p>Connectivity had quietly become essential infrastructure.</p>
<div class="quote">
Access to technology is becoming a fundamental human right.
</div>
<p>In the AI era this question becomes even more urgent.</p>
<hr>
<h2>Enter the AI first world</h2>
<p>Martha believes the technology landscape has shifted once again.</p>
<div class="quote">
We're now living in an AI first world.
</div>
<p>Artificial intelligence tools now appear at extraordinary speed. Systems released a few months ago already feel outdated as new capabilities arrive.</p>
<p>The shift is profound because it changes who can create technology.</p>
<p>For decades software development required specialist skills and expensive infrastructure. Today many tools are accessible to anyone curious enough to experiment.</p>
<p>Which leads to Martha’s most striking observation.</p>
<hr>
<h2>The price of initiative has collapsed</h2>
<div class="pullquote pullquote-right">
The price of initiative has collapsed.
</div>
<p>To illustrate the point Martha described a simple weekend experiment.</p>
<p>She enjoys hosting dinner parties and spends far too long thinking about seating arrangements. With a few AI tools she built a small application that helped organise the seating plan.</p>
<p>It took less than an hour.</p>
<p>No large engineering team. No long development cycle. Just curiosity and accessible tools.</p>
<p>If that kind of experimentation becomes normal, creativity expands dramatically.</p>
<p>The barriers to building ideas are lower than at any moment in the history of technology.</p>
<hr>
<h2>The new divide</h2>
<p>Yet the same tools that create opportunity can also create inequality.</p>
<p>Martha sees a growing gap between two groups.</p>
<p>Those who understand what these tools can do.</p>
<p>And those who have never tried them.</p>
<div class="quote">
The real difference will be between people who understand what is possible and people who don’t.
</div>
<p>Access matters.</p>
<p>That means ensuring AI tools appear not just in research labs and technology companies, but in schools, job centres, small businesses and local communities.</p>
<p>Learning happens through use.</p>
<p>Curiosity grows through experimentation.</p>
<hr>
<h2>The skills that will matter most</h2>
<p>When asked about the skills needed for the future workforce, Martha offered an answer that surprised some people.</p>
<p>She studied classics and history rather than computer science.</p>
<p>Her view is that the most important capabilities are not purely technical.</p>
<p>Critical thinking.</p>
<p>Curiosity.</p>
<p>The ability to ask questions.</p>
<p>Comfort with ambiguity.</p>
<div class="quote">
The most important skills in the future are the ability to ask questions, to interrogate things and to imagine what might be possible.
</div>
<p>Technical expertise will remain vital. Engineers and researchers continue to shape the foundations of modern technology.</p>
<p>Yet in a world where tools evolve rapidly, adaptability becomes equally important.</p>
<hr>
<h2>The importance of diversity in building systems</h2>
<p>Another theme Martha returned to repeatedly was representation.</p>
<p>Technology shapes society. The people building these systems influence the outcomes they produce.</p>
<p>Yet many technology organisations still lack diversity in leadership and decision making roles.</p>
<p>That absence carries economic consequences.</p>
<p>Products built by narrow teams risk missing entire groups of users.</p>
<div class="pullquote pullquote-right">
Hundreds of billions are wasted when we fail to include diverse voices in the workforce.
</div>
<p>More diverse teams create stronger ideas, better products and healthier technology ecosystems.</p>
<hr>
<h2>A moment for the UK technology sector</h2>
<p>Martha also challenged the UK audience to recognise something positive.</p>
<p>Britain remains one of the most significant technology economies in the world.</p>
<p>The country produces world class research.</p>
<p>Investment continues to flow into AI companies.</p>
<p>Startups and scaleups are emerging across sectors.</p>
<p>The challenge is not capability.</p>
<p>The challenge is confidence.</p>
<p>The UK already plays a major role in shaping global technology. Recognising that position helps build momentum for what comes next.</p>
<hr>
<h2>Choosing optimism</h2>
<p>Martha closed her talk with a message that balanced realism with optimism.</p>
<p>She has lived through several waves of technological transformation.</p>
<p>The early internet boom.</p>
<p>The dot com crash.</p>
<p>The rise of mobile platforms.</p>
<p>The expansion of digital services.</p>
<p>Now the AI era.</p>
<p>Each wave created disruption. Each wave also created opportunity.</p>
<div class="quote">
I want to tilt towards the light.
</div>
<p>The AI era offers extraordinary possibilities. Creativity is expanding. Barriers to building ideas are falling. Entire industries are beginning to rethink how work happens.</p>
<p>If access widens, if diverse voices shape the systems we build, and if organisations invest seriously in skills, the next phase of the digital economy could unlock remarkable progress.</p>
<p>For the next generation entering the workforce, the tools available will be more powerful than anything previous generations imagined.</p>
<p>Which makes this moment less about fear and more about responsibility.</p>
<p>The future of work is being written right now.</p>

<h2>Agentic orchestration is the discipline of coordinating multiple AI agents, tools, and data systems so they can complete complex tasks reliably. Instead of one model answering one question, orchestration manages a full workflow: planning, delegation, execution, validation, escalation, and audit.</h2>
<p>If that sounds familiar, it should. The progression looks a lot like how autonomy evolves in cars.</p>
<p>BMW and the wider automotive industry describe automation as a maturity curve from assisted driving to full autonomy. The same framing is useful for enterprise agent systems, where the question is not "does it work once?" but "can it operate safely at scale?"<sup><a href="#user-content-fn-bmw-road" id="user-content-fnref-bmw-road" data-footnote-ref="" aria-describedby="footnote-label">1</a></sup><sup><a href="#user-content-fn-sae-j3016" id="user-content-fnref-sae-j3016" data-footnote-ref="" aria-describedby="footnote-label">2</a></sup></p>
<h2>What Agentic Orchestration Actually Does</h2>
<p>At a practical level, orchestration handles five core jobs:</p>
<ol>
<li><strong>Task decomposition</strong>: breaking a business request into smaller executable steps</li>
<li><strong>Role assignment</strong>: selecting which agent or service should do each step</li>
<li><strong>Control flow</strong>: deciding sequence, parallelism, retries, and fallbacks</li>
<li><strong>Guardrails</strong>: enforcing policy, permissions, compliance, and quality thresholds</li>
<li><strong>Observability</strong>: logging state transitions, outputs, decisions, and exceptions</li>
</ol>
<p>Without orchestration, agent systems often become fragile demos. With orchestration, they become operational systems.</p>
<h2>A Simple Comparison Model: BMW's Five Levels and Agentic Maturity</h2>
<p>BMW's driverless automation narrative is often presented in five levels (from assisted to fully autonomous). For AI teams, this is a useful analogy.<sup><a href="#user-content-fn-bmw-road" id="user-content-fnref-bmw-road-2" data-footnote-ref="" aria-describedby="footnote-label">1</a></sup><sup><a href="#user-content-fn-bmw-five-steps" id="user-content-fnref-bmw-five-steps" data-footnote-ref="" aria-describedby="footnote-label">3</a></sup></p>









































<table><thead><tr><th>BMW automation level</th><th>Vehicle behavior</th><th>Agentic orchestration equivalent</th><th>Human role</th></tr></thead><tbody><tr><td><strong>Level 1: Driver Assistance</strong></td><td>Car can assist steering or speed in limited situations</td><td>Single agent assists a user in one step (for example, draft summary + one tool call)</td><td>Human leads every decision</td></tr><tr><td><strong>Level 2: Partial Automation</strong></td><td>Car can control steering and speed, but driver constantly supervises</td><td>Multi-step workflows execute automatically, but every critical action needs explicit approval</td><td>Human supervises continuously</td></tr><tr><td><strong>Level 3: Conditional Automation</strong></td><td>Car drives in specific conditions and requests takeover when needed</td><td>Agents run end-to-end in defined domains with confidence checks and escalation triggers</td><td>Human intervenes on exceptions</td></tr><tr><td><strong>Level 4: High Automation</strong></td><td>Car drives itself in geofenced domains without constant driver attention</td><td>Orchestrated multi-agent systems run autonomously in bounded business contexts with policy engines and runtime governance</td><td>Human monitors outcomes, not each action</td></tr><tr><td><strong>Level 5: Full Automation</strong></td><td>Car operates everywhere without driver intervention</td><td>Fully autonomous cross-domain orchestration with self-optimizing planning, robust memory, and universal policy compliance</td><td>Human sets goals and strategy only</td></tr></tbody></table>
<p>The level descriptions above align with BMW's published five-level explainer and the SAE J3016 automation taxonomy used across the industry.<sup><a href="#user-content-fn-bmw-road" id="user-content-fnref-bmw-road-3" data-footnote-ref="" aria-describedby="footnote-label">1</a></sup><sup><a href="#user-content-fn-sae-chart" id="user-content-fnref-sae-chart" data-footnote-ref="" aria-describedby="footnote-label">4</a></sup></p>
<h2>Why This Analogy Helps Teams</h2>
<p>Many organizations incorrectly jump from Level 1 prototypes to Level 4 expectations. They expect autonomous operations from systems that still need constant prompting and manual correction.</p>
<p>Using a staged maturity model helps teams:</p>
<ul>
<li>set realistic rollout expectations</li>
<li>design governance proportional to autonomy</li>
<li>align architecture choices with actual risk tolerance</li>
<li>measure progress with clear operational milestones</li>
</ul>
<p>In short, it turns vague "agent strategy" talk into an engineering roadmap.</p>
<div class="quote">Reliable autonomy is not achieved by smarter prompts alone. It is achieved by orchestration, guardrails, and operational discipline.</div>
<h2>Technical Markers for Each Stage</h2>
<p>If you want to know your current stage, check for these markers:</p>
<ul>
<li><strong>Level 1-2 posture</strong>: prompt chaining, ad hoc tool use, low policy enforcement, weak tracing</li>
<li><strong>Level 3 posture</strong>: workflow engine, deterministic checkpoints, approval gates, structured exception handling</li>
<li><strong>Level 4 posture</strong>: policy-as-code, continuous evaluation, memory/state management, observability tied to SLAs</li>
<li><strong>Level 5 posture</strong>: adaptive planning across domains, formal verification for safety-critical actions, autonomous optimization loops</li>
</ul>
<p>Most enterprise deployments today are between Level 2 and Level 3.</p>
<h2>The Core Insight</h2>
<p>Autonomy is not a model feature. It is a systems property.</p>
<p>BMW did not reach higher levels of vehicle automation by adding only better sensors; it required integrated control systems, safety engineering, fallback logic, and regulatory alignment. Agentic orchestration follows the same pattern.<sup><a href="#user-content-fn-bmw-road" id="user-content-fnref-bmw-road-4" data-footnote-ref="" aria-describedby="footnote-label">1</a></sup><sup><a href="#user-content-fn-nhtsa-levels" id="user-content-fnref-nhtsa-levels" data-footnote-ref="" aria-describedby="footnote-label">5</a></sup></p>
<p>If your architecture cannot explain who made a decision, why it was made, what constraints were applied, and how to recover when things fail, you do not yet have high autonomy. You have automation theater.</p>
<h2>Where to Start</h2>
<p>A pragmatic starting path for most teams is:</p>
<ol>
<li>Define one narrow domain where failure impact is controlled</li>
<li>Implement orchestration with explicit state transitions and audit logs</li>
<li>Add policy gates before any external side effects</li>
<li>Introduce confidence-based escalation to humans</li>
<li>Expand scope only after reliability and governance metrics are stable</li>
</ol>
<p>That path will not feel flashy, but it is how reliable autonomy is built.</p>
<p>The same lesson from driverless systems applies directly to AI systems: maturity beats hype.</p>
<section data-footnotes="" class="footnotes"><h2 class="sr-only" id="footnote-label">Footnotes</h2>
<ol>
<li id="user-content-fn-bmw-road">
<p>BMW Group. "The Road to Autonomous Driving at BMW Group." <a href="https://www.bmwgroup.com/en/innovation/automated-driving.html" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://www.bmwgroup.com/en/innovation/automated-driving.html"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-bmw-road" data-footnote-backref="" aria-label="Back to reference 1" class="data-footnote-backref">↩</a> <a href="#user-content-fnref-bmw-road-2" data-footnote-backref="" aria-label="Back to reference 1-2" class="data-footnote-backref">↩<sup>2</sup></a> <a href="#user-content-fnref-bmw-road-3" data-footnote-backref="" aria-label="Back to reference 1-3" class="data-footnote-backref">↩<sup>3</sup></a> <a href="#user-content-fnref-bmw-road-4" data-footnote-backref="" aria-label="Back to reference 1-4" class="data-footnote-backref">↩<sup>4</sup></a></p>
</li>
<li id="user-content-fn-sae-j3016">
<p>SAE International. "Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles (J3016)." <a href="https://www.sae.org/standards/content/j3016_202104/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://www.sae.org/standards/content/j3016_202104/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-sae-j3016" data-footnote-backref="" aria-label="Back to reference 2" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-bmw-five-steps">
<p>BMW.com. "Autonomous driving: 5 steps to the self-driving car." <a href="https://www.bmw.com/en/automotive-life/autonomous-driving.html" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://www.bmw.com/en/automotive-life/autonomous-driving.html"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-bmw-five-steps" data-footnote-backref="" aria-label="Back to reference 3" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-sae-chart">
<p>SAE International. "Levels of Driving Automation Visual Chart (J3016)." <a href="https://www.sae.org/binaries/content/assets/cm/content/blog/sae-j3016-visual-chart_5.3.21.pdf" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://www.sae.org/binaries/content/assets/cm/content/blog/sae-j3016-visual-chart_5.3.21.pdf"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-sae-chart" data-footnote-backref="" aria-label="Back to reference 4" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-nhtsa-levels">
<p>NHTSA. "Levels of Automation." <a href="https://www.nhtsa.gov/document/levels-automation" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://www.nhtsa.gov/document/levels-automation"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-nhtsa-levels" data-footnote-backref="" aria-label="Back to reference 5" class="data-footnote-backref">↩</a></p>
</li>
</ol>
</section>

<h2>If you are trying to understand AWS agentic services in 2026, the key is to separate <strong>agent building</strong>, <strong>agent runtime infrastructure</strong>, and <strong>agent governance/data plumbing</strong>.</h2>
<p>At a high level, AWS now offers a coherent stack centered on Amazon Bedrock Agents and Amazon Bedrock AgentCore, with adjacent services for orchestration, retrieval, and safety.<sup><a href="#user-content-fn-bedrock-agents" id="user-content-fnref-bedrock-agents" data-footnote-ref="" aria-describedby="footnote-label">1</a></sup><sup><a href="#user-content-fn-agentcore-overview" id="user-content-fnref-agentcore-overview" data-footnote-ref="" aria-describedby="footnote-label">2</a></sup></p>
<div class="quote">"Amazon Bedrock AgentCore is an agentic platform to build, deploy and operate highly capable agents securely at scale."</div>
<h2>The AWS Agentic Service Landscape (What matters most)</h2>





















































<table><thead><tr><th>Layer</th><th>AWS service/capability</th><th>What it does</th><th>When to use it</th></tr></thead><tbody><tr><td>Agent application layer</td><td><strong>Amazon Bedrock Agents</strong></td><td>Builds agents that plan and execute multi-step tasks using models, APIs, and enterprise data</td><td>You want managed agent orchestration with fast setup</td></tr><tr><td>Multi-agent coordination</td><td><strong>Bedrock Multi-agent Collaboration</strong></td><td>Coordinates specialist agents under a supervisor agent</td><td>You need role-specialized agent teams for complex workflows</td></tr><tr><td>Agent runtime platform</td><td><strong>Amazon Bedrock AgentCore</strong></td><td>Production platform to deploy/operate agents with security, identity, memory, gateway, and observability controls</td><td>You need production-grade operations, not only prototype agents</td></tr><tr><td>Retrieval and grounding</td><td><strong>Amazon Bedrock Knowledge Bases</strong></td><td>Managed RAG with ingestion, retrieval APIs, source attribution, and structured/unstructured support</td><td>Your agents need grounded answers from enterprise data</td></tr><tr><td>Workflow composition</td><td><strong>Amazon Bedrock Flows</strong></td><td>Visual workflow builder linking prompts, agents, knowledge bases, and AWS services</td><td>You want explicit, auditable multi-step GenAI flows</td></tr><tr><td>Safety and policy</td><td><strong>Amazon Bedrock Guardrails</strong></td><td>Configurable safeguards (moderation, PII, prompt-attack protection, grounding checks) across models and workflows</td><td>You need policy-aligned behavior and safer outputs</td></tr><tr><td>Ecosystem procurement</td><td><strong>AI Agents &#x26; Tools in AWS Marketplace</strong></td><td>Catalog to discover and procure partner agents/tools, including MCP/A2A-aware offerings</td><td>You want faster buy/build hybrid delivery</td></tr></tbody></table>
<p>AWS added Bedrock multi-agent collaboration in March 2025 and launched AI agents/tools in AWS Marketplace in July 2025, which accelerated both build and buy paths for agent programs.<sup><a href="#user-content-fn-multi-agent-ga" id="user-content-fnref-multi-agent-ga" data-footnote-ref="" aria-describedby="footnote-label">3</a></sup><sup><a href="#user-content-fn-marketplace-agents" id="user-content-fnref-marketplace-agents" data-footnote-ref="" aria-describedby="footnote-label">4</a></sup></p>
<h2>AgentCore: The new center of gravity for production</h2>
<p>AgentCore entered preview on <strong>July 16, 2025</strong> and became generally available on <strong>October 13, 2025</strong>.<sup><a href="#user-content-fn-agentcore-preview" id="user-content-fnref-agentcore-preview" data-footnote-ref="" aria-describedby="footnote-label">5</a></sup><sup><a href="#user-content-fn-agentcore-ga" id="user-content-fnref-agentcore-ga" data-footnote-ref="" aria-describedby="footnote-label">6</a></sup></p>
<p>AWS positions it as modular services you can adopt independently:</p>
<ul>
<li><strong>Runtime</strong>: serverless execution for agents/tools, including long-running workloads</li>
<li><strong>Gateway</strong>: tool connectivity and discovery, including MCP connectivity and API/Lambda transformation</li>
<li><strong>Policy (Preview)</strong>: real-time policy enforcement integrated with Gateway</li>
<li><strong>Evaluations (Preview)</strong>: live interaction scoring and quality monitoring</li>
<li><strong>Identity</strong>: agent authN/authZ for AWS and third-party systems</li>
<li><strong>Memory</strong>: persistent agent memory handling</li>
<li><strong>Observability</strong>: CloudWatch/OTel-compatible tracing and metrics</li>
<li><strong>Browser</strong>: secure browser runtime for web actions</li>
<li><strong>Code Interpreter</strong>: sandboxed code execution for agent tasks<sup><a href="#user-content-fn-agentcore-overview" id="user-content-fnref-agentcore-overview-2" data-footnote-ref="" aria-describedby="footnote-label">2</a></sup><sup><a href="#user-content-fn-agentcore-ga" id="user-content-fnref-agentcore-ga-2" data-footnote-ref="" aria-describedby="footnote-label">6</a></sup></li>
</ul>
<p>This matters because Bedrock Agents helps you build agent behavior, while AgentCore helps you run that behavior reliably in production.</p>
<h2>How Bedrock Agents and AgentCore fit together</h2>
<p>A practical mental model:</p>
<ol>
<li><strong>Bedrock Agents</strong> defines how the agent reasons, plans, and invokes actions.<sup><a href="#user-content-fn-bedrock-agents" id="user-content-fnref-bedrock-agents-2" data-footnote-ref="" aria-describedby="footnote-label">1</a></sup><sup><a href="#user-content-fn-agents-ga" id="user-content-fnref-agents-ga" data-footnote-ref="" aria-describedby="footnote-label">7</a></sup></li>
<li><strong>Knowledge Bases</strong> provides grounded context and citations for retrieval-heavy tasks.<sup><a href="#user-content-fn-knowledge-bases" id="user-content-fnref-knowledge-bases" data-footnote-ref="" aria-describedby="footnote-label">8</a></sup></li>
<li><strong>Guardrails</strong> enforces safety and policy constraints across model and workflow steps.<sup><a href="#user-content-fn-guardrails" id="user-content-fnref-guardrails" data-footnote-ref="" aria-describedby="footnote-label">9</a></sup></li>
<li><strong>Flows</strong> composes larger multi-step paths around prompts, agents, and services.<sup><a href="#user-content-fn-bedrock-flows" id="user-content-fnref-bedrock-flows" data-footnote-ref="" aria-describedby="footnote-label">10</a></sup></li>
<li><strong>AgentCore</strong> hardens runtime operations (identity, memory, gateway, observability, runtime isolation).<sup><a href="#user-content-fn-agentcore-overview" id="user-content-fnref-agentcore-overview-3" data-footnote-ref="" aria-describedby="footnote-label">2</a></sup><sup><a href="#user-content-fn-agentcore-ga" id="user-content-fnref-agentcore-ga-3" data-footnote-ref="" aria-describedby="footnote-label">6</a></sup></li>
</ol>
<p>If your team is moving from PoC to production, this sequence avoids the common trap of building clever demos without operational controls.</p>
<h2>What is actually "agentic" in AWS today</h2>
<p>If we define "agentic services" as AWS offerings directly intended to build, coordinate, secure, or operate autonomous/semi-autonomous agent workflows, the core list is:</p>
<ol>
<li>Amazon Bedrock Agents</li>
<li>Bedrock Multi-agent Collaboration</li>
<li>Amazon Bedrock AgentCore (Runtime, Gateway, Policy, Evaluations, Identity, Memory, Observability, Browser, Code Interpreter)</li>
<li>Amazon Bedrock Knowledge Bases</li>
<li>Amazon Bedrock Flows</li>
<li>Amazon Bedrock Guardrails</li>
<li>AI Agents and Tools in AWS Marketplace</li>
</ol>
<p>That is the clearest current map for architecture and investment decisions as of <strong>February 26, 2026</strong>.</p>
<h2>Executive summary</h2>
<p>AWS now has a full agent stack, not just a single "agent" feature.</p>
<ul>
<li><strong>Build behavior</strong> with Bedrock Agents</li>
<li><strong>Coordinate specialists</strong> with multi-agent collaboration</li>
<li><strong>Ground responses</strong> with Knowledge Bases</li>
<li><strong>Compose workflows</strong> with Flows</li>
<li><strong>Enforce safety</strong> with Guardrails</li>
<li><strong>Run at scale</strong> with AgentCore</li>
<li><strong>Extend quickly</strong> through Marketplace agent/tool offerings</li>
</ul>
<p>For most enterprises, AgentCore plus Bedrock Agents is the critical boundary between pilot velocity and production reliability.</p>
<section data-footnotes="" class="footnotes"><h2 class="sr-only" id="footnote-label">Footnotes</h2>
<ol>
<li id="user-content-fn-bedrock-agents">
<p>Amazon Bedrock Agents. <a href="https://aws.amazon.com/bedrock/agents/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://aws.amazon.com/bedrock/agents/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-bedrock-agents" data-footnote-backref="" aria-label="Back to reference 1" class="data-footnote-backref">↩</a> <a href="#user-content-fnref-bedrock-agents-2" data-footnote-backref="" aria-label="Back to reference 1-2" class="data-footnote-backref">↩<sup>2</sup></a></p>
</li>
<li id="user-content-fn-agentcore-overview">
<p>Amazon Bedrock AgentCore overview. <a href="https://aws.amazon.com/bedrock/agentcore/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://aws.amazon.com/bedrock/agentcore/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-agentcore-overview" data-footnote-backref="" aria-label="Back to reference 2" class="data-footnote-backref">↩</a> <a href="#user-content-fnref-agentcore-overview-2" data-footnote-backref="" aria-label="Back to reference 2-2" class="data-footnote-backref">↩<sup>2</sup></a> <a href="#user-content-fnref-agentcore-overview-3" data-footnote-backref="" aria-label="Back to reference 2-3" class="data-footnote-backref">↩<sup>3</sup></a></p>
</li>
<li id="user-content-fn-multi-agent-ga">
<p>AWS What's New. "Amazon Bedrock now supports multi-agent collaboration" (Mar 10, 2025). <a href="https://aws.amazon.com/about-aws/whats-new/2025/03/amazon-bedrock-multi-agent-collaboration/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://aws.amazon.com/about-aws/whats-new/2025/03/amazon-bedrock-multi-agent-collaboration/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-multi-agent-ga" data-footnote-backref="" aria-label="Back to reference 3" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-marketplace-agents">
<p>AWS What's New. "Introducing AI agents and tools in AWS Marketplace" (Jul 16, 2025). <a href="https://aws.amazon.com/about-aws/whats-new/2025/07/ai-agents-tools-aws-marketplace/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://aws.amazon.com/about-aws/whats-new/2025/07/ai-agents-tools-aws-marketplace/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-marketplace-agents" data-footnote-backref="" aria-label="Back to reference 4" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-agentcore-preview">
<p>AWS What's New. "Amazon Bedrock AgentCore now available in preview" (Jul 16, 2025). <a href="https://aws.amazon.com/about-aws/whats-new/2025/07/amazon-bedrock-agentcore-preview/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://aws.amazon.com/about-aws/whats-new/2025/07/amazon-bedrock-agentcore-preview/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-agentcore-preview" data-footnote-backref="" aria-label="Back to reference 5" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-agentcore-ga">
<p>AWS What's New. "Amazon Bedrock AgentCore is now generally available" (Oct 13, 2025). <a href="https://aws.amazon.com/about-aws/whats-new/2025/10/amazon-bedrock-agentcore-available/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://aws.amazon.com/about-aws/whats-new/2025/10/amazon-bedrock-agentcore-available/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-agentcore-ga" data-footnote-backref="" aria-label="Back to reference 6" class="data-footnote-backref">↩</a> <a href="#user-content-fnref-agentcore-ga-2" data-footnote-backref="" aria-label="Back to reference 6-2" class="data-footnote-backref">↩<sup>2</sup></a> <a href="#user-content-fnref-agentcore-ga-3" data-footnote-backref="" aria-label="Back to reference 6-3" class="data-footnote-backref">↩<sup>3</sup></a></p>
</li>
<li id="user-content-fn-agents-ga">
<p>AWS News Blog. "Agents for Amazon Bedrock is now available with improved control of orchestration and visibility into reasoning" (Nov 28, 2023). <a href="https://aws.amazon.com/blogs/aws/agents-for-amazon-bedrock-is-now-available-with-improved-control-of-orchestration-and-visibility-into-reasoning/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://aws.amazon.com/blogs/aws/agents-for-amazon-bedrock-is-now-available-with-improved-control-of-orchestration-and-visibility-into-reasoning/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-agents-ga" data-footnote-backref="" aria-label="Back to reference 7" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-knowledge-bases">
<p>Amazon Bedrock Knowledge Bases. <a href="https://aws.amazon.com/bedrock/knowledge-bases/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://aws.amazon.com/bedrock/knowledge-bases/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-knowledge-bases" data-footnote-backref="" aria-label="Back to reference 8" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-guardrails">
<p>Amazon Bedrock Guardrails. <a href="https://aws.amazon.com/bedrock/guardrails/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://aws.amazon.com/bedrock/guardrails/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-guardrails" data-footnote-backref="" aria-label="Back to reference 9" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-bedrock-flows">
<p>AWS What's New. "Amazon Bedrock Flows is now generally available with two new capabilities" (Nov 22, 2024). <a href="https://aws.amazon.com/about-aws/whats-new/2024/11/amazon-bedrock-flows-new-capabilities/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://aws.amazon.com/about-aws/whats-new/2024/11/amazon-bedrock-flows-new-capabilities/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-bedrock-flows" data-footnote-backref="" aria-label="Back to reference 10" class="data-footnote-backref">↩</a></p>
</li>
</ol>
</section>

<h2>If you are building AI agents in production, the OWASP Top 10 for LLMs and GenAI is not optional reading. It is one of the clearest, community-driven security baselines for how modern LLM and agentic systems fail, and how teams should reduce those failures in real deployments.<sup><a href="#user-content-fn-owasp-project" id="user-content-fnref-owasp-project" data-footnote-ref="" aria-describedby="footnote-label">1</a></sup><sup><a href="#user-content-fn-owasp-llm-top10" id="user-content-fnref-owasp-llm-top10" data-footnote-ref="" aria-describedby="footnote-label">2</a></sup></h2>
<p>OWASP now positions this work under the broader GenAI Security Project, explicitly covering LLMs, agentic AI systems, and AI-driven applications. That scope makes it directly relevant for agent architects, platform teams, security leaders, and engineering executives making delivery and governance decisions.<sup><a href="#user-content-fn-owasp-project" id="user-content-fnref-owasp-project-2" data-footnote-ref="" aria-describedby="footnote-label">1</a></sup><sup><a href="#user-content-fn-owasp-flagship" id="user-content-fnref-owasp-flagship" data-footnote-ref="" aria-describedby="footnote-label">3</a></sup></p>
<h2>Executive Summary</h2>
<p>The 2025 OWASP Top 10 identifies ten recurring risk classes that show up across the LLM and GenAI application lifecycle: development, deployment, and operations.<sup><a href="#user-content-fn-owasp-llm-top10" id="user-content-fnref-owasp-llm-top10-2" data-footnote-ref="" aria-describedby="footnote-label">2</a></sup></p>
<p>The main executive takeaway is simple: most critical failures in agent systems are not model-IQ problems. They are systems-security and control-plane problems.</p>
<p>When translated to agent development, the OWASP list says that teams must secure:</p>
<ul>
<li>inputs and instructions (prompt injection, system prompt leakage)</li>
<li>data and model lineage (supply chain, data/model poisoning)</li>
<li>tool execution boundaries (improper output handling, excessive agency)</li>
<li>retrieval and memory surfaces (vector and embedding weaknesses)</li>
<li>decision quality and operational resilience (misinformation, unbounded consumption)</li>
<li>confidentiality controls (sensitive information disclosure)</li>
</ul>
<p>In short, safe agent development is less about one strong prompt and more about layered controls around planning, tool use, memory, identity, validation, and monitoring.</p>
<div class="quote">This is critical for anyone building AI agents to understand before shipping to production.</div>
<h2>OWASP 2025 Top 10 Through an Agent Development Lens</h2>




























































<table><thead><tr><th>OWASP risk</th><th>What it means in practice</th><th>Why agent teams should care</th></tr></thead><tbody><tr><td><strong>LLM01: Prompt Injection</strong></td><td>Malicious or crafted input manipulates model behavior</td><td>Agents can be pushed into unsafe tool calls, policy bypass, or hidden instruction following</td></tr><tr><td><strong>LLM02: Sensitive Information Disclosure</strong></td><td>Confidential data leaks through prompts, context, or outputs</td><td>Agents often touch internal knowledge, APIs, and user data, increasing leak impact</td></tr><tr><td><strong>LLM03: Supply Chain</strong></td><td>Compromised models, datasets, plugins, or dependencies</td><td>Agent stacks depend on many third-party components and connectors</td></tr><tr><td><strong>LLM04: Data and Model Poisoning</strong></td><td>Training/fine-tuning/embedding data is tampered with</td><td>Poisoned knowledge can silently alter agent behavior and trustworthiness</td></tr><tr><td><strong>LLM05: Improper Output Handling</strong></td><td>LLM output is consumed without validation/sanitization</td><td>Unsafe outputs can become executable actions in downstream systems</td></tr><tr><td><strong>LLM06: Excessive Agency</strong></td><td>Agents get too much autonomy or privilege</td><td>Over-scoped permissions turn normal mistakes into high-impact incidents</td></tr><tr><td><strong>LLM07: System Prompt Leakage</strong></td><td>Hidden instructions are exposed and exploited</td><td>Revealed policy logic makes bypass attempts easier and more targeted</td></tr><tr><td><strong>LLM08: Vector and Embedding Weaknesses</strong></td><td>Weaknesses in retrieval/indexing pipelines</td><td>RAG-enabled agents can be manipulated through poisoned or adversarial retrieval content</td></tr><tr><td><strong>LLM09: Misinformation</strong></td><td>Confident but incorrect outputs influence decisions</td><td>Agentic automation can propagate bad decisions at machine speed</td></tr><tr><td><strong>LLM10: Unbounded Consumption</strong></td><td>Uncontrolled usage of tokens/compute/tooling</td><td>Agent loops and broad tool access can create cost spikes and service instability</td></tr></tbody></table>
<p>Risk names and descriptions above are based on the OWASP 2025 Top 10 content and linked risk pages.<sup><a href="#user-content-fn-owasp-llm-top10" id="user-content-fnref-owasp-llm-top10-3" data-footnote-ref="" aria-describedby="footnote-label">2</a></sup><sup><a href="#user-content-fn-owasp-llm01" id="user-content-fnref-owasp-llm01" data-footnote-ref="" aria-describedby="footnote-label">4</a></sup></p>
<h2>Why This Is Essential Reading for Agent Development</h2>
<ol>
<li>It gives a shared language across engineering, security, and leadership for AI risk prioritization.</li>
<li>It reframes agent failures as architecture and governance issues, not only model behavior issues.</li>
<li>It aligns security planning to the full lifecycle (build, deploy, operate), which is where most agent incidents emerge.</li>
<li>It is directly connected to OWASP's broader agentic security work, making it practical for teams moving from LLM apps to autonomous agents.<sup><a href="#user-content-fn-owasp-project" id="user-content-fnref-owasp-project-3" data-footnote-ref="" aria-describedby="footnote-label">1</a></sup><sup><a href="#user-content-fn-owasp-asi" id="user-content-fnref-owasp-asi" data-footnote-ref="" aria-describedby="footnote-label">5</a></sup></li>
</ol>
<h2>What to Do Next (Executive Action)</h2>
<ul>
<li>Adopt the OWASP Top 10 as a baseline control framework for all new agent initiatives.</li>
<li>Map each agent workflow to the Top 10 and identify where controls are currently missing.</li>
<li>Prioritize guardrails around tool invocation, privilege boundaries, output validation, and retrieval integrity first.</li>
<li>Add continuous monitoring for prompt attacks, data leakage, cost anomalies, and unsafe autonomous actions.</li>
<li>Re-review architecture decisions whenever agent scope or autonomy increases.</li>
</ul>
<p>Agent development is now a security-critical discipline. Teams that treat this OWASP guidance as foundational design input, rather than post-launch compliance, will ship faster with fewer high-severity failures.</p>
<section data-footnotes="" class="footnotes"><h2 class="sr-only" id="footnote-label">Footnotes</h2>
<ol>
<li id="user-content-fn-owasp-project">
<p>OWASP Foundation. "OWASP Top 10 for Large Language Model Applications." <a href="https://owasp.org/www-project-top-10-for-large-language-model-applications/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://owasp.org/www-project-top-10-for-large-language-model-applications/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-owasp-project" data-footnote-backref="" aria-label="Back to reference 1" class="data-footnote-backref">↩</a> <a href="#user-content-fnref-owasp-project-2" data-footnote-backref="" aria-label="Back to reference 1-2" class="data-footnote-backref">↩<sup>2</sup></a> <a href="#user-content-fnref-owasp-project-3" data-footnote-backref="" aria-label="Back to reference 1-3" class="data-footnote-backref">↩<sup>3</sup></a></p>
</li>
<li id="user-content-fn-owasp-llm-top10">
<p>OWASP GenAI Security Project. "2025 Top 10 Risk &#x26; Mitigations for LLMs and Gen AI Apps." <a href="https://genai.owasp.org/llm-top-10/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://genai.owasp.org/llm-top-10/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-owasp-llm-top10" data-footnote-backref="" aria-label="Back to reference 2" class="data-footnote-backref">↩</a> <a href="#user-content-fnref-owasp-llm-top10-2" data-footnote-backref="" aria-label="Back to reference 2-2" class="data-footnote-backref">↩<sup>2</sup></a> <a href="#user-content-fnref-owasp-llm-top10-3" data-footnote-backref="" aria-label="Back to reference 2-3" class="data-footnote-backref">↩<sup>3</sup></a></p>
</li>
<li id="user-content-fn-owasp-flagship">
<p>OWASP GenAI Security Project. "OWASP Top 10 for LLM is now the GenAI Security Project and promoted to OWASP Flagship status" (March 26, 2025). <a href="https://genai.owasp.org/2025/03/26/project-owasp-promotes-genai-security-project-to-flagship-status/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://genai.owasp.org/2025/03/26/project-owasp-promotes-genai-security-project-to-flagship-status/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-owasp-flagship" data-footnote-backref="" aria-label="Back to reference 3" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-owasp-llm01">
<p>OWASP GenAI Security Project. "LLM01:2025 Prompt Injection." <a href="https://genai.owasp.org/llmrisk/llm01-prompt-injection/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://genai.owasp.org/llmrisk/llm01-prompt-injection/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-owasp-llm01" data-footnote-backref="" aria-label="Back to reference 4" class="data-footnote-backref">↩</a></p>
</li>
<li id="user-content-fn-owasp-asi">
<p>OWASP GenAI Security Project. "OWASP Agentic Security Initiative (ASI)." <a href="https://owasp.org/www-project-top-10-for-large-language-model-applications/initiatives/agent_security_initiative/" target="_blank" rel="noopener noreferrer" class="footnote-url-link" title="https://owasp.org/www-project-top-10-for-large-language-model-applications/initiatives/agent_security_initiative/"><svg xmlns="http://www.w3.org/2000/svg" width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" style="display:inline;vertical-align:middle"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></a> <a href="#user-content-fnref-owasp-asi" data-footnote-backref="" aria-label="Back to reference 5" class="data-footnote-backref">↩</a></p>
</li>
</ol>
</section>

<h2>Most enterprise conversations about data still orbit analytics, dashboards, and reporting. Even recent excitement around AI has largely reframed data as something to retrieve and summarise rather than something that actively shapes behaviour. Agent native data challenges that framing.</h2>
<p>The term is intentionally provocative. It suggests that data is no longer designed primarily for people to read, analyse, or debate in meetings that could probably have been emails. Instead, it is designed for software agents that observe environments, reason over current conditions, and take actions inside real workflows.</p>
<div class="pullquote pullquote-right">
Agent native data is what happens when you stop asking what humans want to see and start asking what autonomous systems must know in order to act safely.
</div>
<p>This shift has implications at business, executive, and technical levels, and it explains why so many agent initiatives quietly fail when built on data platforms optimised for yesterday’s problems.</p>
<h2>A concise definition</h2>
<p>Agent native data is data that is stateful, actionable, machine interpretable, and policy aware. It exists to support autonomous or semi autonomous decision making systems, not retrospective analysis.</p>
<h2>How agent native data differs from what came before</h2>
<p>Traditional enterprise data has been shaped by business intelligence and analytics. RAG oriented data has been shaped by document retrieval for large language models. Agent native data is shaped by execution.</p>















































<table><thead><tr><th>Dimension</th><th>Traditional analytics data</th><th>RAG oriented data</th><th>Agent native data</th></tr></thead><tbody><tr><td>Primary purpose</td><td>Reporting and trends</td><td>Knowledge lookup</td><td>Decision making and action</td></tr><tr><td>Typical structure</td><td>Tables, logs, metrics</td><td>Documents, chunks, embeddings</td><td>Structured state, events, constraints</td></tr><tr><td>Time sensitivity</td><td>Batch or near real time</td><td>Mostly static</td><td>Real time and transactional</td></tr><tr><td>Mutability</td><td>Periodic updates</td><td>Read mostly</td><td>Continuously updated by actions</td></tr><tr><td>Semantics</td><td>Human interpreted</td><td>LLM interpreted</td><td>Machine executable</td></tr><tr><td>Governance</td><td>Process driven</td><td>Content filters</td><td>Policy enforced at data layer</td></tr></tbody></table>
<p>The key distinction is that agent native data is not just observed. It is acted upon, updated, and relied on as a source of operational truth.</p>
<h2>Autonomy demands discipline</h2>
<p>Executives often ask whether agents can be trusted. The uncomfortable truth is that trust does not come from better prompts or larger models. It comes from disciplined data foundations.</p>
<p>Agent native data embeds constraints directly into the operational layer. Permissions, entitlements, and policies are checked before actions occur, not after incidents are reviewed.</p>
<h2>The technical view: designed for observe reason act loops</h2>
<p>Technically, agent native data aligns with how agents actually operate. Agents run <strong>observe reason act</strong> loops. They do not continuously reread documents. They observe state changes, reason over structured inputs, and invoke tools.</p>
<p>This drives several design characteristics.</p>
<details>
<summary>Stateful by design</summary>
<div class="accordion-content">
<p>Agents need durable memory of what has happened and what is currently true. This is typically modelled using relational databases, key value stores, or graph models. Vector databases are useful for knowledge retrieval but are ill suited for representing workflow state or commitments.</p>
</div>
</details>
<details>
<summary>Event driven and observable</summary>
<div class="accordion-content">
<p>Agents react to change. Event streams, change data capture, and message buses allow agents to observe transitions as they happen. Querying static snapshots is not enough.</p>
</div>
</details>
<details>
<summary>Action compatible semantics</summary>
<div class="accordion-content">
<p>Agent native data encodes valid states and valid transitions. Preconditions and postconditions are explicit. This allows agents to validate actions before execution and avoid illegal state changes.</p>
<p>In healthcare, for example, a bed is not just free or occupied. It moves through defined operational states, and not all transitions are allowed.</p>
</div>
</details>
<details>
<summary>Policy embedded at the data layer</summary>
<div class="accordion-content">
<p>Because agents can act, access control cannot be bolted on later. Attribute based access control, purpose limitation, and full lineage are integral. This is where many experimental agent systems quietly fail in production environments.</p>
</div>
</details>
<details>
<summary>Optimised for tools, not prompts</summary>
<div class="accordion-content">
<p>Agent native data is accessed through APIs and tools that return typed objects. Language models reason over structured summaries rather than raw data dumps. This aligns naturally with MCP style tool contracts and agent to agent interoperability, where structure matters more than eloquence.</p>
</div>
</details>
<h2>An architectural interpretation</h2>
<p>In practice, agent native data spans three architectural planes.</p>
<div class="grey-box">
<p><strong>The operational state plane</strong> contains workflow state, task queues, short term memory, and transactional facts. This is where decisions are grounded.</p>
</div>
<div class="grey-box">
<p><strong>The knowledge plane</strong> contains policies, rules, reference data, and documents. This supports reasoning but does not drive execution on its own.</p>
</div>
<div class="grey-box">
<p><strong>The coordination plane</strong> contains events, inter agent messages, traces, and telemetry. This enables safe collaboration between agents.</p>
</div>
<p>Most current AI platforms over invest in the knowledge plane. Agent native systems succeed by strengthening the operational state and coordination planes first.</p>
<hr>
<h2>References and further reading</h2>
<p>ReAct Synergizing Reasoning and Acting in Language Models<br>
<a href="https://arxiv.org/abs/2210.03629">https://arxiv.org/abs/2210.03629</a></p>
<p>LangGraph durable execution and state management<br>
<a href="https://langchain-ai.github.io/langgraph/">https://langchain-ai.github.io/langgraph/</a></p>
<p>Event Driven Architecture by Martin Fowler<br>
<a href="https://martinfowler.com/articles/201701-event-driven.html">https://martinfowler.com/articles/201701-event-driven.html</a></p>

<p>Vector databases have become almost synonymous with LLM memory. That association is understandable. Semantic search is powerful, and embeddings underpin modern retrieval augmented generation. However, once we move beyond chatbots and into agentic systems that operate real workflows, the assumption that all memory must live in a vector store starts to unravel.</p>
<p>The underlying issue is simple but often overlooked. Memory and knowledge are different problems, and they benefit from very different storage and retrieval strategies.</p>
<h2>Memory and Knowledge Solve Different Jobs</h2>
<p>Knowledge retrieval focuses on finding relevant information by meaning, typically from unstructured sources such as policies, research papers, or historical conversations. Memory, in contrast, is about preserving facts, state, preferences, and outcomes over time so that future decisions are consistent, safe, and correct.</p>
<div class="pullquote pullquote-right">
“We reached for embeddings out of habit, not because the agent actually needed them.”
</div>
<p>Most of what operational agents need to remember is highly structured. Examples include whether consent has been captured, which workflow step is currently active, which actions have already been taken, or which constraints apply to a specific user, patient, or case. These are not fuzzy concepts. They are deterministic facts.</p>
<p>That distinction becomes clearer when viewed side by side.</p>



































<table><thead><tr><th>Aspect</th><th>Knowledge Retrieval</th><th>Agent Memory</th></tr></thead><tbody><tr><td>Primary goal</td><td>Find relevant text by meaning</td><td>Preserve facts and state over time</td></tr><tr><td>Data type</td><td>Unstructured or semi structured text</td><td>Structured or strongly typed data</td></tr><tr><td>Typical queries</td><td>“Find similar guidance”</td><td>“What is the current status?”</td></tr><tr><td>Tolerance for approximation</td><td>High</td><td>Very low</td></tr><tr><td>Compliance suitability</td><td>Limited</td><td>High</td></tr></tbody></table>
<p>When an agent needs to know whether a patient has been transferred to theatre, or whether a customer has already been contacted, cosine similarity is not just unnecessary. It introduces risk. In these cases the correct operation is a primary key lookup or a filtered query, not a probabilistic match.</p>
<h2>Where Vector Stores Actually Add Value</h2>
<p>Vector databases excel when an agent needs to reason over large volumes of narrative content. Common examples include searching clinical notes, locating relevant policies, recalling earlier conversations by topic, or discovering related incidents across a corpus.</p>
<p>Problems emerge when teams attempt to use vector stores as a general purpose memory layer. This often leads to higher cost, increased latency, weaker explainability, and more complex data governance, without improving decision quality for most operational tasks.</p>
<p>In mature agent architectures, vector memory is therefore complementary rather than foundational. The dominant pattern is hybrid memory, where structured systems hold authoritative state and vector systems provide contextual enrichment when semantic reasoning is genuinely required.</p>
<h2>What Usually Works Better Than Embeddings for Memory</h2>
<p>For most long term agent memory, conventional data stores are not only sufficient, they are preferable.</p>






























<table><thead><tr><th>Storage type</th><th>Why it works well for memory</th><th>Typical uses in agents</th></tr></thead><tbody><tr><td>Relational databases (Postgres, MySQL)</td><td>Strong consistency, transactions, joins, auditability</td><td>Profiles, workflow state, approvals, patient records</td></tr><tr><td>Document databases (MongoDB, DynamoDB)</td><td>Flexible schemas, JSON native storage</td><td>Session memory, evolving agent state models</td></tr><tr><td>Key value stores (Redis)</td><td>Extremely fast access, ideal for checkpoints</td><td>Conversation state, resumable workflows</td></tr><tr><td>Vector databases</td><td>Semantic similarity over text</td><td>Notes, policies, summaries, long narratives</td></tr></tbody></table>
<p>Relational databases are particularly well suited to operational agents because they align with how enterprises already manage critical data. They support transactional safety, deterministic queries, and established governance models, which are essential in regulated environments such as healthcare and finance.</p>
<p>Document databases are useful where agent memory structures evolve rapidly or differ by agent type. Key value stores are commonly used for checkpointing long running workflows, allowing execution to pause and resume safely without reconstructing state.</p>
<p>None of these require embeddings. In many cases, embeddings actively complicate the design.</p>
<h2>How Agents Actually Retrieve Memory</h2>
<p>Modern agent frameworks do not treat memory as passive context that is blindly injected into prompts. Instead, memory is typically accessed through explicit retrieval steps in the agent graph, often implemented as tools or dedicated nodes.</p>
<p>This allows the agent to reason about whether memory is required, which memory is relevant, and how that information should influence the next action. Retrieval may occur before execution begins to hydrate the graph with known state, or during execution when a decision requires additional context.</p>
<p>From the agent’s perspective, retrieving memory via a SQL query, an API call, or a key value lookup is no different from calling a vector search endpoint. What matters is that the retrieval mechanism matches the nature of the data and the decision being made.</p>
<p>For most operational decisions, deterministic queries are faster, cheaper, easier to explain, and easier to govern.</p>
<h2>Why This Matters in Regulated and Operational Domains</h2>
<p>In healthcare, public sector, and financial services, memory must be traceable, correct, and explainable. If an agent recommends an action, it must be possible to demonstrate which facts were used and why.</p>
<p>Vector similarity can suggest that two documents are related, but it cannot establish causality or guarantee factual correctness. That makes it unsuitable as the primary store of authoritative memory in environments where auditability is mandatory.</p>
<p>This is why production agent architectures increasingly separate concerns. Structured systems manage truth and state, while vector systems provide supporting context when narrative reasoning is required. This approach also aligns with emerging interoperability patterns such as the Model Context Protocol, where memory and tools are externalised as services rather than embedded inside prompts.</p>
<h2>TLDR;</h2>
<p>Vector databases are excellent for semantic knowledge retrieval, but they are not a universal memory layer for agents.</p>
<p>If an agent needs to remember facts, preferences, workflow state, or prior decisions, conventional databases are almost always the right starting point. They are cheaper, faster, easier to govern, and better aligned with enterprise and clinical operating models.</p>
<p>Vector memory should be added deliberately, and only where semantic recall over unstructured text genuinely improves reasoning. In well designed agent systems, memory is treated as data first and embeddings second.</p>
<hr>
<h2>References</h2>
<p>LangGraph Persistence and Checkpointing Concepts — LangChain<br>
<a href="https://langchain-ai.github.io/langgraph/concepts/persistence/">https://langchain-ai.github.io/langgraph/concepts/persistence/</a></p>
<p>LangChain RAG and Vector Retrieval Patterns<br>
<a href="https://python.langchain.com/docs/use_cases/question_answering/">https://python.langchain.com/docs/use_cases/question_answering/</a></p>
<p>Model Context Protocol (MCP) — Externalised Context and Tools<br>
<a href="https://modelcontextprotocol.io/">https://modelcontextprotocol.io/</a></p>
<p>Anthropic Research: Agent Architectures and Tool Use<br>
<a href="https://www.anthropic.com/engineering/building-effective-agents">https://www.anthropic.com/engineering/building-effective-agents</a></p>

<h2>Enterprise integration is entering a new phase. For more than a decade, API gateways have been the practical centre of gravity for exposing services, enforcing security policies, and observing traffic across microservices and partner APIs. That model still matters, but the “client” on the other side of the gateway is changing. Increasingly, the caller is not a mobile app or a web front end. It is an AI agent planning a multi step task, invoking tools, streaming results, and delegating subtasks to other agents.</h2>
<div class="pullquote pullquote-right">
API gateways taught us how to govern traffic. Agent gateways will teach us how to govern decisions. That shift is not about AI experimentation, it is about operational control at machine scale.
</div>
<p>That shift is exactly why the agent gateway concept is gaining momentum: it extends the governance and control strengths of API gateways into agent to tool and agent to agent interactions, including emerging protocols such as Model Context Protocol and Agent to Agent.</p>
<h2>What changes when the caller is an agent</h2>
<p>A traditional gateway is optimised for stateless request response patterns: validate credentials, apply rate limits, route to a backend, log outcomes, repeat. An agent workflow is different in three important ways:</p>
<ol>
<li>
<p><strong>Stateful, multi step interactions</strong><br>
Agents operate across sessions, often chaining tool calls with accumulated context, streaming responses, and intermediate reasoning. Gateways must therefore understand sessions and workflows, not just individual calls.</p>
</li>
<li>
<p><strong>Protocol native tool calling and agent messaging</strong><br>
Agent gateways increasingly natively support MCP for tool invocation and A2A for agent coordination, brokering these interactions under central governance rather than leaving them embedded in application code.</p>
</li>
<li>
<p><strong>A larger governance surface area</strong><br>
Governance now extends beyond API access to include which agent can call which tool, under what policy, cost limits, and safety constraints, with full auditability of behaviour.</p>
</li>
</ol>
<h2>The capability stack that defines an agent gateway</h2>








































<table><thead><tr><th>Capability area</th><th>API gateway default</th><th>Agent gateway expectation</th></tr></thead><tbody><tr><td>Protocol handling</td><td>HTTP, REST, gRPC</td><td>MCP tool calls, A2A messaging, protocol mediation</td></tr><tr><td>Identity and access</td><td>OAuth, JWT, API keys</td><td>Machine identity, zero trust agent and tool boundaries</td></tr><tr><td>Traffic controls</td><td>Rate limits, quotas</td><td>Token limits, semantic caching, budget enforcement</td></tr><tr><td>Discovery</td><td>Static endpoints</td><td>Dynamic tool and agent catalogues</td></tr><tr><td>Observability</td><td>Request metrics</td><td>Step level traces, cost analytics, audit replay</td></tr><tr><td>Orchestration</td><td>Limited</td><td>Workflow mediation and tool sequencing</td></tr></tbody></table>
<p>The point is not that all of this must be deployed at once. It is that the architectural boundary for governance is shifting towards where agent actions occur.</p>
<h2>How the major vendors are signalling the shift</h2>
<h3>Google Apigee: enterprise ready AI gateway capabilities</h3>
<p>Google positions Apigee as an AI ready API management platform, adding model abstraction, multicloud model routing, request and response enrichment, semantic caching, token monitoring, and agent connectivity. These features are designed to help organisations control cost, latency, and risk as agents consume and orchestrate AI services.</p>
<h3>Kong: MCP governance and production tooling</h3>
<p>Kong has been explicit about supporting MCP within its AI Gateway, including MCP based tool discovery via its developer portal, alongside observability and governance for tools, prompts, and workflows. Kong also places strong emphasis on machine identity for AI agents and autonomous systems, aligning agent security with existing zero trust patterns.</p>
<h3>Gravitee: agent management as a platform layer</h3>
<p>Gravitee frames the transition as AI agent management, covering agent gateways, agent catalogues, and tool servers. It highlights governance of agent to agent communication using A2A, as well as authentication, rate limiting, monitoring, and discovery as first class concerns for agent ecosystems.</p>
<h3>Tyk: governance agents and distributed synchronisation</h3>
<p>Tyk takes a distributed approach, using deployable governance agents that synchronise API and policy inventories back to a central hub over secure channels. This reflects real world hybrid estates and supports a gradual evolution towards agent aware governance without assuming a single centralised runtime.</p>
<h2>What executives and architects should focus on</h2>
<p>Four themes consistently emerge when organisations move beyond experimentation:</p>
<ol>
<li>
<p><strong>Tool boundary design and permissions</strong><br>
Per tool RBAC and clear scoping are essential to prevent uncontrolled agent behaviour.</p>
</li>
<li>
<p><strong>Cost governance that can be enforced</strong><br>
Token budgets, caching strategies, and routing policies must be part of integration architecture, not just model configuration.</p>
</li>
<li>
<p><strong>Traceability for risk and compliance</strong><br>
Multi step audit trails and replayable execution histories are required when agents influence operational or financial outcomes.</p>
</li>
<li>
<p><strong>Alignment with platform engineering</strong><br>
Policies, configurations, and agent gateway behaviour need to integrate cleanly with infrastructure as code and DevSecOps practices.</p>
</li>
</ol>
<p>If there is one lesson learned early, it is that retrofitting governance after agents are in production is far harder than establishing it upfront.</p>
<h2>Questions to ask your API gateway experts</h2>
<details>
<summary>1. How do you govern agent to tool and agent to agent interactions at runtime?</summary>
<div class="accordion-content">
<p>Agent support is not about exposing tools, it is about controlling behaviour during execution. Without runtime governance, multi step agent workflows quickly become opaque and risky.</p>
<p><strong>Listen for</strong></p>
<ul>
<li>Native support for MCP and A2A with runtime policy enforcement</li>
<li>Per agent and per tool authorisation that persists across sessions</li>
</ul>
</div>
</details>
<details>
<summary>2. How do you provide observability and auditability across multi step agent workflows?</summary>
<div class="accordion-content">
<p>Once agents reason and act across multiple steps, traditional API logging is insufficient. Organisations need visibility that supports operational assurance, investigation, and compliance.</p>
<p><strong>Listen for</strong></p>
<ul>
<li>End to end tracing across reasoning, tool calls, and responses</li>
<li>Replayable execution and audit trails</li>
</ul>
</div>
</details>
<details>
<summary>3. How do you enforce cost, usage, and safety limits for agents in production? </summary>
<div class="accordion-content">
Unbounded agent execution creates immediate financial and operational risk. Cost and safety controls must be enforced at runtime, not left to application logic.
<p><strong>Listen for</strong></p>
<ul>
<li>Token and budget enforcement per agent and per tool</li>
<li>Prompt and response filtering with policy driven guardrails</li>
</ul>
</div>
</details>
<details>
<summary>4. How does your agent support integrate with existing API, DevSecOps, and platform engineering practices?</summary>
<div class="accordion-content">
Agent governance cannot operate as a parallel control plane. If it does not integrate with existing delivery, security, and identity models, it will not scale beyond pilots.
<p><strong>Listen for</strong></p>
<ul>
<li>Infrastructure as code and GitOps compatibility</li>
<li>Alignment with existing API gateway and IAM investments</li>
</ul>
</div>
</details>
<p>Most organisations will evolve their existing API gateways rather than replace them. The practical move is to assess which agent gateway capabilities are required now, such as MCP support, machine identity, cost controls, and deep observability, and which can follow later. Framing agent gateways as the next evolution of API management keeps control with architecture and platform teams, while still enabling rapid agent adoption.</p>
<hr>
<h2>References</h2>
<ol>
<li>
<p>Google Cloud – <em>Apigee AI and API Management</em><br>
<a href="https://cloud.google.com/solutions/apigee-ai">https://cloud.google.com/solutions/apigee-ai</a></p>
</li>
<li>
<p>Kong – <em>Press releases and AI Gateway announcements</em><br>
<a href="https://konghq.com/company/press-room/press-release">https://konghq.com/company/press-room/press-release</a></p>
</li>
<li>
<p>Gravitee – <em>AI Agent Management Platform</em><br>
<a href="https://www.gravitee.io/platform/ai-agent-management">https://www.gravitee.io/platform/ai-agent-management</a></p>
</li>
<li>
<p>Tyk – <em>Agent Management and Governance</em><br>
<a href="https://tyk.io/docs/tyk-governance/agent-management">https://tyk.io/docs/tyk-governance/agent-management</a></p>
</li>
</ol>

<h2>Are we at the beginning of a "software defined care" era, or is this simply another wave of digital health optimism? The question matters because something fundamental is shifting in how care is planned, delivered, and improved. This shift is not about replacing clinicians or reducing healthcare to code. It is about software becoming the primary coordination layer for care decisions, operational flow, and learning at scale.</h2>
<p>"Software defined care" describes a model where clinical and operational decisions are increasingly shaped, triggered, and supported by software systems that continuously learn from vast bodies of evidence and real world data. In this model, software does not just record activity after the fact. It actively participates in how care is delivered, how workflows adapt, and how resources are allocated in real time.</p>
<p>For decades, healthcare software has focused on documentation, billing, and reporting. The new frontier is decisional and operational software. This includes agentic services that can assess context, weigh options, and recommend or trigger actions. It includes automation that removes administrative friction. It includes workflows that adapt dynamically rather than forcing staff to work around rigid systems.</p>
<p>A key driver of this shift is the maturity of large language models that have been trained on millions of healthcare references, research papers, clinical guidelines, and case studies. These models bring a form of statistical memory that no individual clinician can match. They do not replace judgment, but they dramatically expand the accessible knowledge base at the point of care.</p>
<div class="quote">
Consider routine care pathways. Medication reconciliation, standard post operative monitoring, chronic disease follow ups, discharge planning, and routine diagnostics follow patterns that are well understood and heavily documented. These pathways consume a large proportion of healthcare capacity, yet they often rely on manual coordination and human memory. Software defined care targets this space first, where consistency and scale matter most.
</div>
<p>Agentic services play a central role. An agentic system can monitor patient context, detect when predefined criteria are met, and initiate appropriate actions. This might include proposing reflex lab orders when certain clinical thresholds are crossed, alerting teams to emerging risks, or coordinating downstream tasks across departments. The system does not act blindly. It acts within guardrails defined by clinical governance and policy.</p>
<p>Administrative automation is another major component. Prior authorisations, referrals, appointment scheduling, discharge paperwork, and follow up communications consume enormous amounts of staff time. Automating these processes does not reduce care quality. It increases the time available for direct patient interaction. In a software defined care model, administration becomes a background process rather than a dominant workload.</p>
<p>Workflow automation extends this further. Instead of static pathways, workflows become responsive. A delay in imaging can automatically re sequence downstream activities. A change in patient status can trigger a revised care plan. Capacity constraints can be surfaced early rather than discovered too late. Software becomes the conductor that keeps the system in rhythm.</p>
<p>Reflex orders for labs and diagnostics illustrate the efficiency gains clearly. In many settings, clinicians already follow well established rules for when certain tests should be ordered. Encoding these rules into software reduces delay and variation. It also ensures that routine decisions are applied consistently, while still allowing clinicians to intervene when nuance is required.</p>
<div class="grey-box">The distinction between routine and complex care is critical. Software defined care delivers its greatest gains in routine treatments where evidence is strong and variation adds little value. Complex cases, rare conditions, and ethically sensitive decisions still demand deep human expertise, empathy, and shared decision making. The goal is not to automate everything, but to free human intelligence for the situations where it matters most.</div>
<p>This raises an uncomfortable but necessary question. In the future, will more decisions be shaped by a doctor with ten years of experience, or by an LLM trained on ten million books and case studies? The honest answer is that it will be both. The clinician brings context, accountability, and human judgment. The model brings breadth, recall, and pattern recognition at a scale no human can achieve. Software defined care is about combining these strengths rather than choosing between them.</p>
<h2>Three cornerstones underpin this emerging model.</h2>
<h3>The first cornerstone is evidence at scale.</h3>
<p>Decisions are informed by continuously updated bodies of knowledge that include guidelines, outcomes data, and real world practice patterns. Software ensures that this evidence is available at the moment decisions are made, not buried in journals or static protocols.</p>
<h3>The second cornerstone is agentic orchestration.</h3>
<p>Care delivery is coordinated by intelligent services that understand goals, constraints, and dependencies. These agents do not replace teams. They support them by managing complexity, sequencing actions, and highlighting trade offs in real time.</p>
<h3>The third cornerstone is human governance and trust.</h3>
<p>Software defined care only works if clinicians trust the systems and retain meaningful control. Transparency, auditability, and explainability are essential. Decisions must be understandable, challengeable, and aligned with professional accountability.</p>
<h2>Software defined care is not a distant future concept.</h2>
<p>Elements of it are already appearing across healthcare systems, often quietly and incrementally. The real question is not whether it will arrive, but how intentionally it will be designed. If done well, it offers a path to more consistent, efficient, and humane care by allowing software to handle the routine and predictable, while people focus on the complex and profoundly human aspects of medicine.</p>

<p>January 2026 marks a clear inflection point for artificial intelligence in healthcare. Over the course of just a few weeks, four major developments revealed how large language models and agentic AI are moving from broad experimentation into domains where trust, safety, and measurable impact matter deeply. Together, these updates tell a connected story about maturity. They show an industry learning where AI fits, where it does not, and how it must be governed when human health is involved.</p>
<div class="quote">Rather than a single breakthrough, this moment reflects a transition. AI is no longer only about impressive demonstrations. It is about accountability, regulation, and integration into clinical and scientific workflows.</div>
<h2>When General AI Meets Health Risk</h2>
<p>On <strong>2 January 2026</strong>, an investigation by <em>The Guardian</em> brought renewed scrutiny to AI generated health information appearing in consumer search results. Google had been rolling out AI Overviews that summarised answers at the top of search pages. While many worked as intended, several health related summaries were found to contain misleading or incomplete medical advice.</p>
<p>Examples included incorrect interpretations of blood test ranges and oversimplified nutritional guidance that lacked clinical context. Medical professionals warned that such summaries could be harmful when detached from patient specific factors such as age, sex, medical history, or medication use. In response, Google removed a number of these health related AI summaries and stated that improvements were underway to reduce the risk of harm.</p>
<p>This moment matters because it highlights a fundamental truth. General purpose AI systems, when placed in front of millions of people, can influence health decisions even when they were not designed to act as medical tools. The issue was not malicious intent, but misplaced authority. A short summary presented with confidence can feel definitive, even when medicine rarely is.</p>
<p>The takeaway was not that AI has no role in health, but that <strong>context and safeguards are essential</strong>. Without them, scale amplifies error.</p>
<h2>A Shift Toward Regulated Clinical AI</h2>
<p>Just days later, OpenAI announced the launch of <strong>ChatGPT for Healthcare</strong>, reframing how LLMs are positioned in medical environments. Unlike consumer chat tools, this offering is designed explicitly for hospitals, clinics, and healthcare organisations operating under strict regulatory frameworks.</p>
<p>Crucially, the platform is <strong>HIPAA compliant</strong>, supporting encryption, audit logs, access controls, and Business Associate Agreements. These features acknowledge a reality long understood by healthcare providers. Innovation without compliance is not adoption.</p>
<p>The intent behind ChatGPT for Healthcare is not to diagnose patients or replace clinicians. Instead, it focuses on supporting workflows that already exist. Examples include summarising clinical notes, assisting with documentation, helping staff navigate internal policies, and synthesising large volumes of structured and unstructured information.</p>
<p>This distinction matters. Where consumer AI struggled by offering answers without sufficient guardrails, this model is embedded within governance. The system is constrained by design, monitored by organisations, and operated by trained professionals. It reflects a growing understanding that healthcare AI must be <strong>enterprise first</strong>, not consumer first.</p>
<p>The contrast with the earlier search controversy is striking. One highlights the risks of uncontextualised medical information at scale. The other shows what happens when AI is deployed inside regulated boundaries with accountability built in.</p>
<h2>From Language to Discovery in the Lab</h2>
<p>While clinical workflows were evolving, another development expanded the scope of healthcare AI far beyond text. In <strong>January 2026</strong>, NVIDIA and Eli Lilly announced an expanded partnership focused on AI driven drug discovery. Their collaboration centres on using advanced AI platforms to accelerate the identification and development of new therapies.</p>
<p>Drug discovery is a slow and expensive process. Traditional pipelines can take over a decade and cost billions. The partnership aims to change that by applying AI systems that can analyse biological data, simulate molecular interactions, and guide experimental decisions. This is where agentic AI becomes especially relevant.</p>
<p>Unlike simple prediction models, agentic systems can manage sequences of tasks. They can evaluate results, propose next experiments, and optimise workflows over time. In this context, AI does not just analyse data. It participates in the research process itself.</p>
<p>The significance here is scale. This is not a pilot or proof of concept. It is a strategic investment by a major pharmaceutical company and one of the world’s leading AI infrastructure providers. It signals confidence that AI is ready to influence the earliest stages of medicine, long before a patient ever enters a clinic.</p>
<p>It also shows how healthcare AI is diversifying. Language models support people. Agentic systems support discovery. Both are needed, but they operate in very different risk profiles and timelines.</p>
<p><img src="/blog/llm.jpg" alt="alt"></p>
<h2>Seeing and Hearing Medicine with MedGemma</h2>
<p>Completing this picture, Google Research released <strong>MedGemma 1.5</strong> in January 2026, extending AI capabilities into medical imaging and clinical speech. This update builds on earlier MedGemma models by improving interpretation of medical images and introducing advanced medical speech to text functionality.</p>
<p>MedGemma 1.5 reflects a broader shift toward multimodal AI. Healthcare data is rarely just text. It includes scans, images, spoken notes, and signals. Bringing these together allows AI systems to understand clinical situations more holistically.</p>
<p>In practical terms, this means radiology images can be analysed alongside dictated observations, or spoken clinical notes can be transcribed and structured with medical awareness. Early benchmarks suggest strong performance in specific imaging tasks, pointing toward AI that can meaningfully assist specialists rather than simply automate clerical work.</p>
<p>Importantly, MedGemma is positioned as a research and development tool, not a standalone diagnostic authority. This reinforces a pattern seen across all four developments. The most credible progress comes when AI is framed as an assistant within expert led systems.</p>
<h2>An Emerging Pattern of Maturity</h2>
<p>Viewed together, these January updates reveal how healthcare AI is growing up.</p>
<p>The removal of misleading search summaries shows the limits of general AI when applied without context. The launch of a HIPAA compliant clinical platform demonstrates how those limits can be addressed through governance and design. The NVIDIA and Lilly partnership illustrates AI’s role in accelerating scientific discovery where human experimentation alone struggles with scale. MedGemma 1.5 highlights the move toward multimodal systems that reflect how medicine actually works.</p>
<p>What connects these stories is not technology alone, but <strong>intent</strong>. The industry is learning to ask better questions. Where should AI speak directly to patients, if at all. Where should it support clinicians behind the scenes. Where can it safely explore possibilities in laboratories and research environments.</p>
<p>This moment does not signal an end state. It signals a recalibration. LLMs and agentic AI are no longer novelties in healthcare. They are tools being shaped by regulation, ethics, and real world constraints. The progress of January 2026 shows that the future of healthcare AI will be defined less by bold claims and more by careful integration into systems that already carry profound responsibility.</p>

<p>The future of enterprise AI isn't about isolated models making decisions in silos. It's about intelligent agents that communicate, collaborate, and orchestrate complex workflows across your entire technology stack. Today, we're exploring two transformative integration patterns that are reshaping how we build agentic systems: <strong>Agent-to-Agent (A2A) communication</strong> and the <strong>Model Context Protocol (MCP)</strong>.</p>
<blockquote>
<p><strong>Explore the code</strong>: <a href="https://github.com/riverthink/agenticintegration">github.com/riverthink/agenticintegration</a></p>
</blockquote>
<p>These aren't just theoretical concepts. We've built <a href="https://github.com/riverthink/agenticintegration">working demonstrations</a> that showcase exactly how these patterns unlock unprecedented capabilities for healthcare providers, retail operations, and financial institutions. Let's dive into what makes these approaches so powerful.</p>
<h2>The Power of Connected Intelligence</h2>
<p>Imagine a world where your AI agents don't just respond to queries—they actively collaborate to solve complex problems. Where a patient-facing healthcare agent can seamlessly delegate to a specialized medical records agent. Where a retail customer service agent can consult with inventory, pricing, and shipping agents in real-time. Where financial advisors have AI assistants that orchestrate data from compliance, market analysis, and portfolio management systems.</p>
<p>This is the promise of modern agentic architecture, and it's available today.</p>
<h2>Demo 1: Agent-to-Agent Communication with A2A Protocol</h2>
<h3>The Architecture</h3>
<p>Our <a href="https://github.com/riverthink/agenticintegration/tree/main/a2a-example">A2A demonstration</a> showcases how two independent agents can communicate using a standardized protocol. Here's what makes it special:</p>
<p><a href="https://github.com/riverthink/agenticintegration/blob/main/a2a-example/agent_a.py"><strong>Agent A</strong></a> runs on port 9999 and acts as an orchestrator. When it receives a request, it doesn't try to handle everything itself. Instead, it delegates to <a href="https://github.com/riverthink/agenticintegration/blob/main/a2a-example/agent_b.py"><strong>Agent B</strong></a> (running on port 9998), which specializes in processing specific types of requests.</p>
<p>Both agents leverage <strong>LangGraph</strong>, a framework for building stateful workflows, combined with the <strong>A2A SDK</strong> for standardized agent communication. The result? A clean, scalable architecture where agents can discover each other, communicate via HTTP, and collaborate on complex tasks.</p>
<h3>How It Works</h3>
<p>The implementation is elegantly simple yet powerful:</p>
<ol>
<li><strong>Agent B</strong> sets up a LangGraph workflow that processes incoming messages and returns structured responses</li>
<li><strong>Agent A</strong> maintains its own LangGraph workflow that calls Agent B via the A2A client</li>
<li>Messages flow through the A2A protocol with proper task tracking and event queuing</li>
<li>Each agent exposes an <strong>AgentCard</strong> that describes its capabilities, skills, and communication preferences</li>
<li>The system handles streaming responses, cancellation, and error recovery automatically</li>
</ol>
<p>The beauty of this pattern lies in its modularity. Agent B doesn't need to know anything about Agent A beyond the protocol specification. Agent A can delegate to multiple specialized agents without managing complex integration code.</p>
<h3>Real-World Applications</h3>
<p><strong>Healthcare: Coordinated Patient Care</strong></p>
<p>Picture a hospital system where different AI agents specialize in different aspects of patient care:</p>
<ul>
<li>A triage agent receives patient symptoms and delegates to diagnostic specialists</li>
<li>A diagnostic agent consults with medical history, lab results, and imaging agents</li>
<li>Treatment planning agents coordinate with pharmacy, scheduling, and insurance verification agents</li>
</ul>
<p>Each agent maintains its specialized knowledge while communicating through the A2A protocol. The result? Faster, more accurate care coordination without overwhelming human clinicians.</p>
<p><strong>Retail: Intelligent Customer Experience</strong></p>
<p>Retail operations involve countless interconnected decisions:</p>
<ul>
<li>A customer inquiry agent delegates to inventory checkers, pricing agents, and delivery estimators</li>
<li>A returns processing agent coordinates with warehouse systems, refund processors, and customer satisfaction agents</li>
<li>A personalization agent consults purchase history, preference analyzers, and recommendation engines</li>
</ul>
<p>By decomposing these complex workflows into specialized agents, retailers can build systems that are both more powerful and easier to maintain.</p>
<p><strong>Finance: Orchestrated Decision-Making</strong></p>
<p>Financial services demand accuracy, compliance, and speed:</p>
<ul>
<li>A client-facing advisor agent delegates to market analysis, risk assessment, and compliance checking agents</li>
<li>Trading agents coordinate with data feeds, model executors, and regulatory reporting agents</li>
<li>Fraud detection agents collaborate with transaction analyzers, pattern recognizers, and case management agents</li>
</ul>
<p>The A2A pattern enables financial institutions to build sophisticated AI systems while maintaining clear audit trails and separation of concerns.</p>
<h2>Demo 2: Model Context Protocol for Tool Integration</h2>
<h3>The Architecture</h3>
<p>While A2A handles agent-to-agent communication, MCP solves a different problem: How do we give agents access to external tools and data sources in a standardized way?</p>
<p>Our <a href="https://github.com/riverthink/agenticintegration/tree/main/mcp-example">MCP demonstration</a> features:</p>
<ul>
<li>An <a href="https://github.com/riverthink/agenticintegration/blob/main/mcp-example/mcp_sse_server.py"><strong>MCP Server</strong></a> that exposes weather and traffic data tools via Server-Sent Events (SSE)</li>
<li>An <a href="https://github.com/riverthink/agenticintegration/blob/main/mcp-example/test_mcp_client.py"><strong>MCP Client</strong></a> powered by GPT-4 that intelligently selects and uses the right tools based on user queries</li>
</ul>
<p>The server is built with <strong>FastMCP</strong>, a lightweight framework that makes it trivial to expose Python functions as MCP tools. The client uses LangChain to process natural language queries and make intelligent tool selection decisions.</p>
<h3>How It Works</h3>
<p>The implementation demonstrates the full MCP lifecycle:</p>
<ol>
<li>The server decorates Python functions with <code>@app.tool()</code>, automatically exposing them via the MCP protocol</li>
<li>The client connects to the server's SSE endpoint and retrieves the list of available tools</li>
<li>When a user asks a question, the client sends it to an LLM along with tool specifications</li>
<li>The LLM analyzes the question, selects the appropriate tool, and generates the right arguments</li>
<li>The client calls the selected tool and returns structured results</li>
</ol>
<p>In our demo, asking "What's the weather in Paris?" triggers the LLM to select the <code>get_weather</code> tool with <code>{"city": "Paris"}</code> as arguments. The system handles JSON schema validation, type checking, and error handling automatically.</p>
<h3>Real-World Applications</h3>
<p><strong>Healthcare: Clinical Decision Support</strong></p>
<p>Medical professionals need access to diverse data sources and tools:</p>
<ul>
<li>EHR query tools that retrieve patient histories with natural language</li>
<li>Drug interaction checkers that analyze medication lists</li>
<li>Clinical guideline search tools that find relevant protocols</li>
<li>Lab result interpreters that provide context and flagging</li>
</ul>
<p>An MCP-enabled healthcare agent can access all these tools through a single, standardized interface. Doctors ask questions in natural language, and the agent orchestrates the right combination of tools to provide comprehensive answers.</p>
<p><strong>Retail: Omnichannel Operations</strong></p>
<p>Retail systems connect to countless data sources:</p>
<ul>
<li>Inventory management systems across warehouses and stores</li>
<li>Pricing engines with real-time competitor monitoring</li>
<li>Customer data platforms with purchase history and preferences</li>
<li>Logistics APIs providing shipping estimates and tracking</li>
</ul>
<p>With MCP, a retail agent can query "Show me low-stock items in the Northwest region with high sales velocity" and automatically orchestrate calls to inventory, sales analytics, and geographic segmentation tools.</p>
<p><strong>Finance: Data-Driven Insights</strong></p>
<p>Financial analysis requires integrating multiple data streams:</p>
<ul>
<li>Market data feeds providing real-time prices and indicators</li>
<li>Portfolio management systems with holdings and performance</li>
<li>Risk calculation engines running stress tests and scenarios</li>
<li>Regulatory databases checking compliance requirements</li>
</ul>
<p>MCP enables financial agents to answer complex questions like "What's my portfolio's exposure to tech sector volatility given current market conditions?" by automatically selecting and combining the right analytical tools.</p>
<h2>The Technical Foundation</h2>
<p>Both demonstrations leverage modern Python async patterns, making them highly efficient and scalable:</p>
<p><strong>A2A Demo Stack:</strong></p>
<ul>
<li><strong>A2A SDK</strong>: Standardized protocol for agent discovery and communication</li>
<li><strong>LangGraph</strong>: Stateful workflow orchestration for complex agent behaviors</li>
<li><strong>Uvicorn</strong>: High-performance ASGI server for async HTTP handling</li>
<li><strong>httpx</strong>: Modern async HTTP client for agent-to-agent calls</li>
</ul>
<p><strong>MCP Demo Stack:</strong></p>
<ul>
<li><strong>FastMCP</strong>: Lightweight framework for building MCP servers</li>
<li><strong>LangChain</strong>: LLM integration with tool calling capabilities</li>
<li><strong>OpenAI GPT-4</strong>: Intelligent tool selection and argument generation</li>
<li><strong>Server-Sent Events</strong>: Efficient streaming protocol for tool communication</li>
</ul>
<p>Both patterns are production-ready, with proper error handling, type validation, and extensibility hooks.</p>
<h2>Why These Patterns Matter Now</h2>
<p>We're at an inflection point in enterprise AI adoption. Organizations have moved past proof-of-concept chatbots and are building production agentic systems. But without standardized integration patterns, these systems become brittle, hard to maintain, and impossible to scale.</p>
<p>The A2A protocol and MCP provide the standardization layer that the industry needs. They enable:</p>
<ol>
<li><strong>Composability</strong>: Build specialized agents and combine them in new ways</li>
<li><strong>Interoperability</strong>: Agents from different vendors can communicate using standard protocols</li>
<li><strong>Maintainability</strong>: Changes to one agent don't cascade through the system</li>
<li><strong>Observability</strong>: Standard protocols enable standard monitoring and debugging tools</li>
<li><strong>Security</strong>: Well-defined boundaries make it easier to implement access controls and audit logs</li>
</ol>
<h2>Getting Started</h2>
<p>Both demonstrations are <a href="https://github.com/riverthink/agenticintegration">open source</a> and designed to be educational. You can run them locally in minutes:</p>
<p>The <a href="https://github.com/riverthink/agenticintegration/tree/main/a2a-example">A2A example</a> shows you how to build multi-agent workflows where agents delegate to each other. Start both agents, send a test request, and watch the logs to see how Agent A discovers and communicates with Agent B.</p>
<p>The <a href="https://github.com/riverthink/agenticintegration/tree/main/mcp-example">MCP example</a> demonstrates intelligent tool selection. Start the server, run the client, and ask natural language questions. Watch how the LLM analyzes your question and selects the right tool automatically.</p>
<h2>The Path Forward</h2>
<p>These patterns aren't just demos—they're blueprints for the next generation of intelligent systems. As you explore them, think about:</p>
<ul>
<li>Which parts of your current workflows could benefit from agent delegation?</li>
<li>What data sources and tools could you expose via MCP?</li>
<li>How could specialized agents improve accuracy while reducing complexity?</li>
<li>Where are the natural boundaries between different agent responsibilities?</li>
</ul>
<p>The future of enterprise AI is collaborative, composable, and connected. With patterns like A2A and MCP, that future is being built today.</p>
<h2>Start Building</h2>
<p><a href="https://github.com/riverthink/agenticintegration">Clone the repository</a>, run the examples, and see these patterns in action. Modify the agents to suit your needs. Expose your own tools via MCP. Build agent networks that solve your specific challenges.</p>
<p>The code is clear, <a href="https://github.com/riverthink/agenticintegration#readme">well-documented</a>, and designed for learning. Whether you're in healthcare, retail, finance, or any other industry, these integration patterns provide the foundation for building sophisticated agentic systems.</p>
<p>The question isn't whether your organization will adopt agentic AI. The question is how quickly you'll build the integration infrastructure to make it successful.</p>
<p>Start exploring today. The future is calling—and it's ready to collaborate.</p>
<hr>
<p><strong>About the Demos</strong>: These examples are built with production-quality frameworks and follow best practices for async Python development, error handling, and protocol implementation. They're designed to be both educational and adaptable to real-world use cases.</p>
<p><strong>Try It Yourself</strong>: Both demos include complete setup instructions, test scripts, and sequence diagrams. <a href="https://github.com/riverthink/agenticintegration">Visit the repository</a> to get started, and join the community of developers building the next generation of agentic systems.</p>

Dispatches
from the Edge

Designing Memory Architectures for Agentic AI Systems at Scale

Before You Trust AI in Hospitals, Run This One Type of Trial

Spec-Driven Development: The Power Inversion Your Engineering Team Won't Tell You About

Agent Identity in Healthcare AI - The Control Layer Executives Demand

Google ADK vs LangChain vs LangGraph: The New Stack for Agent Engineering

What Neon Taught Me About AI and Real Skills

Dispatchesfrom the Edge