# Agentic Architecture Skills Author: Murali Sid, Founder and AI Systems Architect. https://linkedin.com/in/muralisid About the author: https://www.agenticarchitectureskills.com/about Licence: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/). Quote freely with attribution to Murali Sid. Cite as: Murali Sid. . Agentic Architecture Skills. # Docs - [Start here: the road to an agentic enterprise](/start): What an agentic enterprise is, the three ideas behind the guide, and the pages to read at each of seven steps, in order. - Use cases - [Explore business opportunities](/use-cases): Pick an industry and see its estate and use cases. Then follow one worked example from today's process to the process with agents. - **Choose an industry** - [Healthcare: follow the work around the patient](/use-cases/healthcare): The fourteen layers of a public hospital network and the systems in each. Four agent use cases, the rules that bind them, and what runs today. - [Energy and resources: the estate and where agents fit](/use-cases/energy): The systems a generator, an oil and gas operator and a miner run, layer by layer. The rule that keeps agents off the control path. The use cases and named deployments placed on that estate. - [Utilities: keep essential services working](/use-cases/utilities): The estate of an electricity distributor and a water utility on the fourteen layers. Four worked examples placed on it, who uses AI today, and the rules that bind the work. - **Worked examples** - [From referral to a confirmed appointment](/use-cases/referral-coordination): A referral to a public hospital clinic, step by step. The systems by layer, the work today, the same work with one agent, and who runs it. - [Keep agreed follow-up from getting lost](/use-cases/care-follow-up): Follow-up after a discharge or a clinic visit, step by step. The systems by layer, the work today, the same work with one agent that tracks evidence, and who closes what. - [Prepare the evidence before a consult or multidisciplinary meeting](/use-cases/consult-briefing): Before a cancer meeting or a clinic, gather results, reports and outside records into one brief. Every fact shows its source and date. The systems by layer, the work today, the same work with one agent, and who signs. - [Investigate an energy asset before planning work](/use-cases/energy-asset-investigation): When plant equipment runs outside its normal range, an agent gathers the evidence from the historian, the asset management system and engineering tools. An engineer decides what work to plan. - [Prioritize vegetation inspections](/use-cases/vegetation-inspection): How an electricity distributor turns imagery into vegetation work today. How an agent can carry a finding to a proposed work request that a planner approves. - [Investigate a water-network anomaly](/use-cases/water-operations): How a water utility's control room investigates an alarm today. How an agent can gather the evidence from a read-only copy of the data while people decide and act. - [Resolve billing exceptions without deciding hardship](/use-cases/billing-exceptions): How an energy retailer clears implausible meter reads and held bills today. How agents can run the same work, with a rule or a specialist approving every change. A water variant is included. - [Safety coaching from CCTV, incident and observation data](/use-cases/safety-coaching): Turn a risky moment caught on camera, or logged as an observation, into coaching the safety owner approves. Real-time interventions, discipline and any decision about a person stay with people. - [Collect and check compliance evidence](/use-cases/compliance-evidence): The assurance pattern that every regulated industry shares, worked through for hospital accreditation against the NSQHS Standards. The systems by layer, the work today, the same work with one agent, and the utility and energy variants. - [Build an AI use inventory a regulator can read](/use-cases/ai-use-inventory): How a utility builds a list of every AI use that a regulator can read. One row per use, with the decision it shapes, its safeguards and the evidence that they work. - [Investigate recurring audit issues across sites](/use-cases/systemic-audit-issues): When the same finding keeps appearing at several sites, an agent tests competing explanations against maintenance, inspection and procedure records. The assurance owner decides the response. - [Investigate production discrepancies](/use-cases/production-measurement): When truck tonnes from the pit and belt-scale tonnes at the plant disagree, an agent checks timing, instruments and material. A named owner accepts any adjustment to the production record. - **Choose and fund** - [Choose where to start](/use-cases/choose-where-to-start): A programme lead has a shortlist and must pick one task for a first agent. Five checks decide which idea goes first and which goes on a preparation list. - [Agents and packaged applications](/use-cases/packaged-applications): Most processes run inside packaged software you cannot rewrite. Two tests decide whether to wait for the vendor's agent, buy it, or build a thin one. - [Measure value and full cost](/use-cases/value-and-investment): A sponsor asks what an agent will cost and what it will save. Count every cost, divide by the outcomes the business accepts as resolved, and compare with today. - [Build an agentic enterprise](/use-cases/enterprise-transformation): One agent works. Grow to many by sharing what repeats, naming owners, counting the agents you buy, and expanding only on measured results. - **Blueprints and frameworks** - Architecture: The cross-layer design of the agentic enterprise. - [Architecture of the agentic enterprise](/architecture): How model intelligence becomes useful enterprise work: the layers you own, the lifecycle every agent runs through, the capabilities you add, and the controls set along the way. - **See the system** - [What an agent platform provides](/architecture/system-view): The seven capabilities that must exist before agents can run safely at scale, the layers that build them, and the lifecycle stages that call on them. - [Follow one inspection task](/architecture/task-walkthrough): See how evidence, reasoning, permission, action and confirmation fit together in one agent task. - **Classify every agent** - [What kind of agent is this? Four questions](/architecture/agent-types): Ask four questions about any agent: what it does for people, what its own credentials let it do, how freely it acts, and what it can damage. The answers set its risk tier and its controls. - [The autonomy contract](/architecture/autonomy-contract): How much an agent may do on its own, and how it may learn, set on two scales for each piece of work. Readiness sets a ceiling, risk can only lower it, and each step up needs controls in place. - [The risk profile: factors, tiers, controls](/architecture/risk-profile): Score any agent on nine factors, each Low, Medium or High. The worst factor sets Tier 1 (basic), Tier 2 (enhanced) or Tier 3 (strict). The tier sets the controls, who signs and how often the agent is reviewed. - **Bound what it can do** - [The four deterministic zones](/architecture/deterministic-zones): The four areas where an AI model may never approve: access, money, safety actuation and formal records. Beside them sit the destruction rule and the rule for decisions about people. - [The delegation chain: trace every action to the person who asked](/architecture/delegation-chain): How every action an agent takes stays traceable to the person who asked, from the agent's own identity, through the gateway, to the system of record. - [The agent lifecycle: register to retire](/architecture/agent-lifecycle): The nine stages every agent passes through, from registration to retirement. A registry entry and a named sponsor come first, before any credential, and retiring is more than switching off. - [Enforcement outside the model](/architecture/enforcement): A rule written in a prompt is a request. The same rule checked by a separate system is a control. Three places enforce outside the model: the gateway, the policy decision point, and fixed rules promoted from what the agent learned. - **Knowledge, learning and reference** - [The data-to-memory pipeline](/architecture/data-to-memory): How company data becomes a cited answer and an agent memory in nine stages, and the rule that every copy keeps the original's access, privacy and deletion duties. - [The learning flywheel](/architecture/learning-flywheel): How an agent's behaviour changes safely: a change passes evaluations, a canary and a staged release, and can be pulled back. - [The concern matrix](/architecture/concern-matrix): Ten requirements every layer must honour, such as identity, privacy and cost. For each one: which layer owns it, which layer stops a violation, and the six gaps with no complete answer yet. - [Make software that agents can use](/architecture/software-for-agents): Give outside agents clear operations, bounded authority and reliable results, so they can use your product without you losing control. - [Plain words](/architecture/plain-words): The glossary: every technical term in this guide in one plain sentence, grouped by theme, with a link to the page that explains it. - Intelligence ladder: Choose the change that improves the task. - [Intelligence ladder](/ladder): How much intelligence does an agent need for its job, and what should that intelligence cost? - [Choose what to change](/ladder/choosing-an-approach): Diagnose the failure before choosing a model or training method. - [Step 1: Instructions](/ladder/instructions): What a prompt is made of, why examples beat adjectives, how a search beats hand-tuning, and when to move on. - [Step 2: Context](/ladder/context): How documents become retrieved evidence, why one vector per document fails, a measured fix, and what an agent should remember. - [Step 3: Tools and the loop](/ladder/tools-and-the-loop): How the agent loop runs, what the harness controls, and when several agents are worth their cost. - [Workflow, one agent, or several?](/ladder/single-and-multiple-agents): Choose coordination that helps the task rather than adding unnecessary moving parts. - [Combine models around a useful task](/ladder/combining-models): Let specialist models contribute evidence while the complete system owns the result. - [Adapters and supervised fine-tuning](/ladder/adapters-and-fine-tuning): How training on examples changes a model, what LoRA and QLoRA are, what it cannot teach, and what it breaks. - [Distillation](/ladder/distillation): How a large model teaches a small one, how it repairs forgetting, and the arithmetic that decides when it pays. - [Reinforcement fine-tuning](/ladder/reinforcement-fine-tuning): How training against a grader works, how graders get gamed, and the narrow tasks where it is the right tool. - [Continued pretraining](/ladder/continued-pretraining): How a model learns a domain from raw text, what preparing that text involves, and where the budget should go. - [Custom pretraining and architecture](/ladder/custom-pretraining): Two meanings of 'our own model': a foundation model built from scratch, or a different model for non-text data. - Knowledge & memory: Remember the work, preserve evidence, and improve future decisions. - [Knowledge & memory](/memory): Help an agent continue its work, use past experience, and support decisions with evidence. - **Start with the foundations** - [What an agent should remember](/memory/memory-types): Working context, thread memory, and durable experiences, facts, and methods. - [From information to usable memory](/memory/information-to-memory): Preserve the evidence while making it easier to find and use. - **Make memory usable** - [Choose how memory is represented](/memory/choosing-representations): Use exact records, similarity, or relationships according to the question. - [From a media match to an event](/memory/images-video-and-audio): Find the relevant interval and preserve what a reviewer needs to check it. - [Bring evidence together](/memory/combining-evidence): Join sources by identity, place, and time before combining their conclusions. - [Keep current facts and history distinct](/memory/changing-facts): Separate event time, arrival time, and the period a fact applies to. - [Control what memory retains and reveals](/memory/permissions-and-lifecycle): Make ownership, access, correction, and removal work across derived information. - [Test memory against the work](/memory/what-the-benchmarks-measure): Distinguish finding evidence, answering correctly, and completing a task. - **Optional: types of data** - [Explore memory in vector space](/memory/vector-spaces): A visual tour of text, image, video, audio, geospatial, and time-series representations. - [Text embeddings: search by meaning](/memory/text-embeddings): Make differently worded passages discoverable while preserving exact evidence. - [Image embeddings: search by visual features](/memory/image-embeddings): Find related inspection images while keeping the original observation available. - [Video embeddings: search for events](/memory/video-embeddings): Represent a useful interval while retaining timing and surrounding context. - [Audio embeddings: search for acoustic resemblance](/memory/audio-embeddings): Keep sound evidence separate from the words a transcript captures. - [Represent places and changes over time](/memory/location-and-satellite-data): Use geospatial similarity while retaining the location and observation period. - [Represent how equipment behaves over time](/memory/time-series): Find similar operating periods without discarding units, duration, or context. - **Advanced design and research** - [Choose how to store memory](/memory/six-architectures): Compare six implementation choices after deciding what the memory is for. - [How context and memory work together](/memory/long-context-is-not-memory): A larger reading window helps some tasks; it does not decide what to retain or trust. - [Token efficiency: spend tokens on outcomes](/memory/token-efficiency): An agent's bill grows with every loop. Measure it per resolved outcome, meter four numbers for each agent, and stop the spend before the call. - [When memory spans years and sites](/memory/the-ten-million-token-question): Scale the search, updates, and evidence checks without losing the question. - [Research findings and what remains open](/memory/what-we-have-measured): Separate author-reported experiments, public evidence, and work still to be reproduced. - [Sources and further reading](/memory/sources): Primary references for the memory story, representations, and evaluation. - Research - [Research and open questions](/research): Explore experiments, their limits, and the findings that change how we design agents. - [Experiments and results](/research/experiments): Read the question, comparison, findings, and limits before applying a result. - [Questions to investigate](/research/questions): Proposed work that could improve the guide, with an explicit comparison and decision to inform. - [What changed our advice](/research/what-changed): Keep the results that challenged the original idea visible. - [Industrial examples from external sources](/research/industrial-examples): What current visual and document systems demonstrate, and what an enterprise agent still needs. - [How we evaluate evidence](/research/method): Make the comparison, limitations, and effect on guidance inspectable. - Skills: Install the guide into your own agent. - [Install this guide into your agent](/skills): Every section of this site is published as an Agent Skill: a folder your agent reads, carrying the architecture, the decisions and the measured evidence, with the sources intact. - [Installing the skills](/skills/install): One section per agent: Claude Code, claude.ai, Codex and ChatGPT, Cursor, Gemini CLI, GitHub Copilot, and installing by hand. All of them install the same bundles. - [How the skills are built](/skills/how-skills-are-built): The bundles are generated from the site's own pages by a build step: how the conversion works, why the archives are byte-reproducible, what the checks enforce, and what the published research says about the limits of skills as a format. - Operate: Run agents day to day, and organise the teams that do. - [Operate: the weekly loop](/operate): The four checks to run every week once an agent is live, the measures the weekly review reads, and the gate before the agent may do more. - [How to organise the teams: One Platform, Three Lanes](/operate/organisation): Who owns the agent platform, who owns each agent and who supervises it: a thin technology office, one platform team, three lanes that build each pattern once and step back, and business teams that own their agents. - Layers: The fourteen enterprise layers, each with its target state, mechanisms, and decisions. - [The fourteen layers](/layers): The fourteen parts of the enterprise estate that agents touch, grouped by the job they do, with one page for each: what sits there today, what agents add, and what must be true first. - [Infrastructure and compute](/layers/r01-infrastructure): Where agents run: a sealed workspace for each session, one guarded exit to the network, an identity for each agent, sessions that survive a crash, and rented model capacity placed by data sensitivity. - [Data platform](/layers/r02-data-platform): How agents get company data they may see and can trust: permission checks before the search, an index protected like the data it copies, fixed definitions for numbers that matter, and freshness matched to the decision. - [Integration fabric](/layers/r03-integration-fabric): How agents reach company systems: one standard way to connect tools, one gateway every request passes through, and the checks, credentials and audit record applied there. - [Systems of record](/layers/r04-systems-of-record): Agents act inside billing, asset, finance and clinical record systems as the person who asked, through the permission checks those systems already run, with risky changes paused for approval. - [Line of business and OT](/layers/r05-lob-and-ot): Agents advise the people who run grids, plants and field work, and never control equipment. Every suggestion is checked in a simulation before display, advice stays apart from alarms, and a person approves every change to the asset register. - [Intelligence and learning](/layers/r06-intelligence-and-learning): How agents get better without quietly getting worse: experts own the tests, model judges grade with human spot checks, and a rule reaches production only after it survives attempts to break it. - [Agent platform](/layers/r07-agent-platform): Where agents run: buy the runtime, build the loop around the model, give every run several ways to stop, keep its state where people can read it, and let only tested versions run. - [Productivity and collaboration](/layers/r08-productivity-and-collaboration): What office tools must provide for agents: every agent gets its own identity with a named sponsor, a mailbox or meeting seat is a rare separate grant, and assistants roll out task by task. - [Experience and channels](/layers/r09-experience-and-channels): What customer channels must provide for agents: a separate customer edge for disclosure, hand-over and evidence, on shared platform capabilities, measured by problems solved rather than calls avoided. - [Security and identity](/layers/r10-security-and-identity): A model cannot tell instructions from data, so build for a tricked agent: its own short-lived identity, a policy check it cannot skip on every action that matters, and filters that warn but never decide. - [Governance, risk and sovereignty](/layers/r11-governance-risk-sovereignty): What the governance layer must provide for agents: one official register with a named sponsor and risk tier for every agent, records that grow with risk, a tested stop, and data that runs only where it may. - [Observability and FinOps](/layers/r12-observability-and-finops): What the observability layer must provide for agents: a record of every run that the agent cannot change, a tested gate on every change, and a budget with hard caps that a named sponsor owns. - [Operating model and supervision](/layers/r13-operating-model): What the supervision layer must provide for agents: a named sponsor for every agent, oversight sized as a queue for the busiest ten minutes, checks made cheap, and measures that prove oversight is happening. - [Agent data engineering](/layers/r14-agent-data-engineering): What the data engineering layer must provide for agents: curated collections with named owners, documents read with their source attached, access rules that travel with every copy, checked memory, and deletion that reaches every copy. - [The Agentic OS: what you buy and what you build](/agentic-os): What an Agentic OS is, the runtime you buy, the harness you build, and the work no Agentic OS does for you. - Security practices: Threat modelling, adversarial evaluation, runtime controls and incident practices for AI agents. - [Security practices for AI agents](/security): An agent reads text an attacker could have written, holds credentials and can act. Four practices keep it useful under attack: threat model, adversarial evaluation, runtime controls and incident practice. - [Threat-model an agentic workflow](/security/threat-model): How to threat-model an agent: add intent, trust, authority, mutable state and effects to the data-flow diagram, derive abuse cases, then bind each one to controls outside the model and a test. - Patterns: Patterns for giving an agent the right context, each with the evidence behind it. - [Context patterns for agents](/patterns): Nine tested ways to put the right material in front of an agent. Each says what to do, when it applies, and the evidence behind it. Patterns one to three give most of the value. - [One vector blurs a document about many topics](/patterns/aspect-dilution): One vector for a document about many topics matches each topic weakly. Several vectors win most of it back. Most of the gain comes from having more pieces, not smarter pieces. - [On real scientific text, plain chunks beat purpose cards](/patterns/real-prose): On 5,183 scientific abstracts judged by people, purpose cards score below plain chunks of the same count. Cards designed by a model do no better. Adding keyword search gives the best result. - [Build the index from the documents, not from the goal](/patterns/why-not-to-condition-on-the-objective): Showing the index designer a sample of the documents helps a lot. Telling it your goal makes search worse, and more context makes it worse still. Tested five ways, ending on real data judged by people. - [Where the cost saving really comes from](/patterns/gate-and-economics): A cheap first-pass filter tells relevant from irrelevant email well, but removes only a third of the documents. The real saving comes from paying the model per topic, not per document. It is about two hundred times cheaper, and the factor does not grow with size. - [Choosing varied results, and whether the reader matters](/patterns/diversity-and-consumer): Keep the most relevant results and fill the rest with the least similar ones. This buys topic coverage at half the relevance cost of the standard method. The idea that a model reader needs more variety than a person did not hold up. - [How to test a context design](/patterns/how-to-test-a-context-design): Give the plain method the same budget as the clever one, and make sure the test can fail. Then write the rules first, let someone try to break the result, and rerun it on a second machine. Five of our own headline results reversed, and two causes explain all five. - [What we got wrong](/patterns/what-we-got-wrong): The scorecard for every prediction in our context experiments. Three held, one held only in a narrower form, and seven were contradicted or not supported. One useful result came by accident. - [The papers behind the patterns](/patterns/reading-list): The published work behind each context pattern, grouped by the question it answers. Each entry says what the paper shows, and whether our results agree, disagree or set a limit. - The decision catalog - [The decision catalog](/decisions): Twenty-five technology choices usually made by habit. Each record gives the common choice, the verdict, the one test that decides your case, and what would change the answer. - [About the guide](/about): What this guide is, who wrote it, where to start, and how its claims are kept honest. ## Agent skills This site publishes its own content as installable Agent Skills, so an agent can carry the architecture, the decisions and the measured evidence without fetching pages. - Discovery index: https://www.agenticarchitectureskills.com/.well-known/agent-skills/index.json - Catalogue and install instructions: https://www.agenticarchitectureskills.com/skills - Install everything: npx skills add muralisid/agentic-architecture-skills Every page is also available as Markdown: append .md to any URL, or send Accept: text/markdown.