Field guide · Free · No email wall

The Agent Development Lifecycle

From prompt to accountable system — how design, evaluation, release, deployment, operations and governance combine into one system.

9 parts 14 figures 10 maps of an agent estate 22 artifacts indexed 24-question readiness check By Emil Krzemiński, auxfirst
Introduction

An agent is much more than a prompt

Looked at from the outside, an AI agent seems to be mostly a prompt. Somebody wrote clever instructions, connected a model, and the thing started working. The prompt matters — but in a mature environment it is one layer out of many.

An agent has to be designed. Someone has to decide what it is allowed to touch. Its tools have to be built, described and versioned. Its context has to be assembled, curated and kept fresh. Its behaviour has to be evaluated, packaged into something you can ship, and promoted into an environment where real consequences live. Then it has to be watched, corrected, secured, upgraded and — eventually — retired.

If an organization runs hundreds of agents across dozens of teams, another problem appears: you need to know what actually exists, who is accountable for it, and what it can reach. That is why the agentic era is generating its own vocabulary of registries, manifests, catalogs, evaluations, ladders and receipts.

Some of it describes behaviour. Some describes permission. Some describes what was deployed. Some exists to answer one deceptively simple question:

If this agent does the wrong thing at 3am, what breaks, who finds out, and can we take it back?

This field guide is a map of that world. It is the agentic successor to the software anatomy every technology organization already knows: SDLC, CI/CD, GitOps, CMDB, SBOM, SLOs, postmortems. Each has an agentic counterpart — sometimes a direct translation, sometimes a concept with no ancestor at all, because software that decides and acts creates problems that software which merely runs never had.

It is not written for a machine-learning researcher. It is written for the technically literate person around agentic systems: the product manager, founder, consultant, investor, operations lead, business architect, marketer or delivery director who has to make decisions about agents without building the runtime themselves.

The point of the guide

The goal is not to learn a hundred names. It is to understand the relationships between them. Where classic software asked what do we have and where does it run, agentic software adds three questions that change everything: what may it do on its own, how do we know it is still behaving, and who answers for the outcome?

Contents · nine parts and a close
  1. IThe five worlds of agentic softwareDevelopment, delivery, deployment, operations, governance — and the ADLC loop
  2. IIDesigning an agent before it existsArchitecture, decision records, the agent stack, inventory and radar
  3. IIIWhere an agent comes fromRepositories, dependencies, the ABOM and the agent supply chain
  4. IVFrom change to productionEvaluation, release, deployment, manifests, GitOps and secrets
  5. VThe units of an agentic systemAgents, tools, contracts, context, provenance and catalogs
  6. VIKnowing what you haveThe agent CMDB, cost, portfolio management and ownership
  7. VIIRunning agents in productionObservability, incidents, service levels, the trust budget, reversal
  8. VIIISecurity, risk and governanceThreat models, credentials, target architecture, policy as code
  9. IXThe graphTen maps of an agent estate, and why agents need them too
  10. Close, artifact index and the 24-point checkWhat each artifact is for, and twenty-four questions worth failing
Part I Contents ↑

The five worlds of agentic software

Chapter 01

Five overlapping worlds

The landscape is easiest to understand split into five overlapping areas.

Figure 1 · The five worlds

Overlapping, not sequential — which is why one artifact is never enough

World 01

Development

Creating the thing. Deciding the job to be delegated, decomposing it into steps, choosing models, writing instructions and skills, building tools, designing what the agent shows a human before it acts.

World 02

Delivery

Turning a working prototype into a version you can safely run. In classic software this is compilation and packaging. Here it is evaluation: proving a specific bundle of model, prompt, tools and policy behaves as intended.

World 03

Deployment

Actually running that version in a specific environment, with specific credentials, against specific systems and real people.

World 04

Operations

Keeping a live agent healthy: watching behaviour, catching drift, reviewing its work, handling escalations, absorbing incidents, controlling cost.

World 05

Management & governance

Running the whole estate: which agents exist, what each may do, who owns it, what data it touches, what it costs, whether it still deserves the autonomy it has.

The overlap between the linear story and the permanent layer below is exactly why an organization ends up needing both an agent manifest and an agent registry, both an evaluation suite and an autonomy map, both a trace and an action receipt. Each looks at the same living system from a different side.
Figure 2 · The linear story, and the layer above it

Every agent runs through this — and is governed across all of it

The path a delegated job takes
Delegated jobAgent designContext & instructionsEvaluationReleaseDeploymentActing agentSupervisionCorrectionNext release
The permanent layer sitting above all of it
InventoryOwnershipPermissionEvidenceObservabilityGovernanceCostCompliance
Chapter 02

ADLC: the map of an agent's whole life

The foundational concept of this guide is the ADLC — Agent Development Lifecycle. It is not a technology or a product. It is a model describing the stages an agent passes through, from the decision to delegate something to the decision to switch it off.

Figure 3 · The ADLC loop

Seven stages, and it does not end

The Agent Development Lifecycle loop Seven stages in sequence — delegate, design, build, evaluate, deploy, supervise, improve — with a return arrow from improve back to design, showing the lifecycle is a loop rather than a line. DELEGATE DESIGN BUILD EVALUATE DEPLOY SUPERVISE IMPROVE EVERY INCIDENT BECOMES A TEST · EVERY TEST BECOMES A CONSTRAINT ON THE NEXT VERSION
Read against the classic SDLC, three differences matter more than the similarities — and each one is highlighted above.
Classic SDLCADLCWhy the shift matters
TestingEvaluationDeterministic software passes or fails a test. An agent produces a distribution of behaviour, so the question moves from "does this return 4" to "across a hundred realistic cases, how often does it do the right thing, how does it fail, and what does it do when unsure". Evaluation is statistical, and it never fully ends.
OperatingSupervisingA deployed service is monitored. A deployed agent is supervised: someone reads a sample of its work, approves the hot actions, and absorbs the escalations. That work has throughput limits — and those limits, not model capability, usually decide how many agents an organization can actually run.
Functional improvementBehavioural improvementA release can change how an agent reasons, refuses, escalates and sounds without changing a single feature.

In older organizations these stages sit in different departments: a data science team builds, an engineering team integrates, a business team uses, and nobody owns the behaviour. The modern answer is the one DevOps gave twenty years ago, shifted one level up. Call it AgentOps:

The AgentOps premise The people who design an agent's behaviour must understand how it is deployed, supervised and reversed.

That small shift produces the whole family of practices in this guide: declarative agent definitions, automated evaluation, behavioural regression suites, action receipts, autonomy maps as living policy, and reversal designed before launch rather than improvised during an incident.

Designing an agent before it exists

Chapter 03

Agent architecture: before there is a prompt

Before anything is built, someone has to decide how the agent will be constructed. Agent architecture describes the significant elements of an agentic system and the relationships between them. It is not the wording of an instruction.

It answers questions like: what job is being delegated, and what does "done" look like? How is that job decomposed into steps the agent can plan over? Which tools does it call, and what can each tool reach? What does it remember, for how long, and who can edit that memory? Which actions may it take alone, which require a human, which are forbidden? Where does it hand off — to a person, a specialist agent, or an external system? And what evidence does it leave behind?

Figure 4 · A straightforward customer-operations agent

Already an architecture diagram — note the three things a classic one never had

Customer messageIntent handshakePlanner
Fans out to
RetrievalCRM toolsRefund tool
Converges on
DraftHuman checkpointAction + receipt
A checkpoint, a receipt, and a moment where the system deliberately slows down to confirm intent — the Intent Handshake pattern. Two diagrams can look identical while one agent drafts emails and the other sends them.
Chapter 04

The agent architecture diagram

There is no single mandatory format, and the level of detail varies enormously. The problem starts when everyone draws it their own way — and in agentic systems that problem is worse than in classic software, because the most important facts are invisible on a normal box-and-arrow drawing.

A useful agent diagram therefore encodes three things a classic diagram never had to:

  • Direction of authority — who delegated to whom, and who is accountable at the end of the chain.
  • Heat — which arrows carry consequence, in the sense of the Action Heat Ladder.
  • Checkpoints — where a human enters, and what happens if they do not respond.
Chapter 05

Zooming an agentic system: the four levels

The C4 model gave software four levels of zoom — context, container, component, code. Agentic systems need the same discipline, with different contents at each level.

Level 1 — Delegation context

Who delegates work to this system, what outside systems does it touch, and what does it produce for whom? An account manager delegates to a campaign agent, which reaches the ad platform, the CRM and the client. At this level nobody cares how the agent is built. The question is: who is trusting it with what, and who sees the result?

Level 2 — Runtimes and stores

The large, separately deployable pieces: the orchestrator, each agent runtime, tool servers (often MCP servers), the memory and vector stores, the evaluation harness, the human review surface.

Level 3 — Components

Inside a runtime: planner, retriever, tool adapters, guardrail and policy checks, the verifier that inspects output before it leaves, the escalation router.

Level 4 — Behavioural source

The lowest level is no longer only code. It is system prompts, skills and instruction files, tool schemas, policy definitions, and the evaluation cases that define what "correct" means.

The idea that survives translation

Good architecture documentation lets you change the level of zoom. Most organizations live comfortably at levels one to three. Level four is where the behaviour actually lives — and where most teams have no documentation at all.

Chapter 06

Agent, system-of-agents, and enterprise architecture

The word architecture covers three very different scopes.

Agent architecture describes one agent: its role, tools, memory, boundaries and escalation paths. The question is how is this agent built?

System-of-agents architecture describes several agents delivering one business outcome together — an orchestrator, specialists, a verifier, humans at defined checkpoints. The question is how do these actors collaborate, and who owns the result? This is the level where attribution becomes hard and where multi-actor clarity stops being a nicety.

Enterprise agent architecture tries to describe the agentic landscape of a whole organization.

Figure 5 · Enterprise agent architecture

The map of the city — including which streets the agents may walk down

Business capabilities
Business processes
Delegated jobs
Agents
Tools & systems
Data & context
Models & infrastructure
If agent architecture is the map of one building, enterprise agent architecture is the map of the city. Read downward it explains dependency; read upward it explains exposure.
Chapter 07

The Agent Decision Record

A diagram shows what exists. It rarely shows why it was designed this way. Classic engineering solved this with the ADR — Architecture Decision Record. Agentic systems need the same artifact, with a wider remit.

Figure 6 · A worked Agent Decision Record

Short, one decision, and it records the reasoning you will forget

DecisionRefunds under €200 are auto-approved by the support agent.
Context92% of refunds are under €200; median human review adds 19 hours.
Options(a) all refunds human-approved  (b) threshold  (c) full autonomy
Choice(b) — threshold with sampled review at 10%
BecauseReversibility is high (refunds can be clawed back within 30 days), blast radius is one customer, commitment is capped.
ConsequenceRequires the refund cap enforced in the tool, not the prompt. Requires weekly sample review by the support lead.
RevisitAfter 500 refunds or any dispute escalation.
Agentic decisions worth recording are not only architectural: autonomy thresholds, escalation recipients, memory retention, tone and refusal policies, model choice, and every place a human was deliberately kept in the loop.

This matters more than in classic software, because agent decisions look arbitrary from the outside. Two years later, a new team sees a confirmation step on a trivial action and removes it as friction — not knowing it was added after a €40,000 incident.

What decision records really are How an organization keeps the memory of its own caution.
Chapter 08

The agent stack

The set of technologies used to build and run an agentic system. Two layers deserve attention because teams routinely forget them until something goes wrong.

Figure 7 · A representative agent stack

Nine layers — and the last two decide whether anyone trusts it

ModelsFrontier model, plus a small fast model for routing
OrchestrationThe loop: planning, tool calls, retries, handoffs
ToolsFunction calling, MCP servers, browser and computer use
ContextRetrieval, instruction files, skills, structured records
MemorySession state, long-term store, user-editable profile
GuardrailsInput filters, output verification, policy engine
EvaluationCase suites, graders, behavioural regression
ObservabilityTraces, transcripts, cost and quality metrics
Human surfaceReview queues, approvals, escape hatches, receipts
Highlighted rows are the two most often deferred — and the two that decide whether the agent is deployable.
Guardrails are not instructions

An instruction is a request; a guardrail is a mechanism the agent cannot talk its way past. If your spending limit lives in the system prompt rather than in the tool, you do not have a limit — you have a preference.

The human surface is equally first-class. Every consequential agent needs somewhere a person can see what it did, disagree, undo and escalate. Teams that treat this as "the UI we'll add later" ship agents nobody is willing to trust with anything meaningful.

Chapter 09

Agent inventory: what we actually have

The naive version is a list of names. A mature one carries the fields that let someone make a decision.

FieldExampleWhy it matters
Agentinvoice-matching agentIdentity, and a name humans recognize
Delegated jobMatch invoices to POs, flag exceptionsBusiness purpose, not technical function
Accountable humanFinance operations leadSomeone answers for outcomes
AutonomyAuto under €2,000; approval aboveThe contract, readable outside the codebase
Hottest actionApprove payment — HIGHRisk is set by the hottest thing it can do
Data reachFinance systems, supplier recordsExposure and privacy surface
Model & versionFrontier model, pinnedUpgrades change behaviour
StatusProduction since MarchDistinguishes pilots from load-bearing systems
Volume & cost~4,000 actions / monthValue, and the review burden it creates

The difference is qualitative. One inventory says "we use an invoice agent". The other says: an agent with payment-approval authority handles 4,000 finance actions a month, owned by one named person, auto-approving below a threshold set nine months ago.

The second is a management artifact. The first is a rumour.
Chapter 10

The agent radar

A list still does not express the organization's stance. The technology-radar model — adopt, trial, assess, hold — translates cleanly, and is unusually valuable here because the ground moves so fast.

Figure 8 · The agent radar

The hold column is the one that earns its keep

Adopt
  • Pinned frontier model
  • MCP for internal tools
  • Eval-gated releases
Trial
  • Computer-use agents in sandbox
  • Agent-to-agent handoffs
Assess
  • Autonomous multi-day workflows
  • Agent payment rails
Hold
  • Unversioned prompts in production
  • Shared admin credentials
In agentic estates the dangerous entries are rarely exotic technologies — they are ordinary shortcuts: prompts edited live in a console, one service account shared by six agents, a pilot that quietly became load-bearing.

An inventory says what you have now. A radar says what you want more of, less of, and what you are watching. Together they turn a moving landscape into something a leadership team can actually steer.

Part IIIContents ↑

Where an agent comes from

Chapter 11

The repository: where an agent lives

Before an agent runs, it has to exist somewhere. In a healthy setup that place is a version-controlled repository — and what lives there is broader than code.

Figure 9 · What is actually in an agent repository

Seven directories, only one of which is code in the classic sense

/promptsSystem prompts, role definitions, refusal policy
/skillsReusable procedures the agent can load on demand
/toolsTool implementations and their schemas
/contextInstruction and knowledge files the agent reads
/policyAutonomy map, limits, escalation rules
/evalsTest cases, golden transcripts, graders
/receiptsSchema of the evidence the agent must emit
Plain-text instruction formats — AGENTS.md, CLAUDE.md, SKILL.md, a business.md describing how a company actually works — have quietly become the substrate agents read. Covered in depth in the Markdown guide and Markdown for AI.
The test

If the agent's behaviour changes, can you point at the commit? If behaviour changes because someone edited a text box in a vendor console, the agent has no source of truth — only a current mood.

Chapter 12

One repository or many

The monorepo/polyrepo decision recurs, with an agentic twist. One repository for the estate makes shared context, shared tools and fleet-wide policy changes trivial — when a refusal rule or an escalation recipient changes, it changes everywhere at once. One repository per agent gives teams separation, independent release cadence, and tighter permissions on sensitive agents.

Neither is universally right. But the agentic version has an extra dimension the software version did not: context inheritance. When agents share a repository they tend to share instruction files — and a change to a shared file silently changes the behaviour of every agent that reads it, including ones whose owners were not in the conversation.

A folder decision quietly becomes a behavioural blast radius. Whichever model you choose, the rule that saves you is the same: shared context files need owners and change review, exactly like shared libraries.

Chapter 13

Dependencies: almost nothing stands alone

Every agent inherits a large surface it did not write: a model with a version, provider and safety settings; a framework or SDK; internal APIs and their permissions; MCP servers, often third-party; retrieval sources including the open web; prompt and skill libraries; and sub-agents, which have dependencies of their own.

Two things make this dependency graph harder than the software one.

First, dependencies arrive through content. A retrieved document, a web page, a tool description written by a vendor — all of these enter the agent's reasoning. A library can only run the code it contains. A retrieved page can contain instructions.

Second, dependencies change without a version bump. A model provider updates a system-level behaviour, a vendor rewrites a tool description, a knowledge base gets a new page. Nothing in your repository changed; your agent behaves differently on Tuesday.

Chapter 14

ABOM: the agent bill of materials

An SBOM lists what is inside a piece of software. An ABOM lists what is inside an agent's behaviour.

Figure 10 · A generated ABOM

Produced at release time, stored with the release — not written by hand

Agentsupport-resolution-agent v4.2
Modelfrontier-model (pinned build) · fallback: small-model
Promptssystem-prompt.md sha 8f31c2…
Skillsrefund-policy.md, tone-guide.md, escalation.md
Toolscrm.read, crm.write, refund.issue (cap €200), email.draft
MCP serversticketing-mcp v1.4 (internal), knowledge-mcp v2.0 (vendor)
Contextproduct-docs snapshot 2026-07-01, policy-base v9
Memoryper-customer profile, 90-day retention, user-editable
Policyautonomy-map v3 · refusal-policy v2
Evalssuite v11 — 240 cases, pass 96.3%, escalation recall 0.98
OwnerHead of Customer Operations
The ABOM turns a vague worry into a two-minute query: a vendor's tool server was compromised — which of our agents used it, with what permissions, and what did they do with it? Without one, that answer takes a week of interviews.

The same artifact also powers model-deprecation planning, licence and data-usage review, and incident scoping. It is the single most under-built artifact in agentic organizations today.

Chapter 15

The agent supply chain

The full chain from intent to action — instructions, context and retrieval, model, plan, tools, action, evidence — can be attacked at every stage. The attacks differ from classic software in one crucial way: the payload is language, not executable code.

  • Instructions can be overridden by text the agent reads later — the essence of prompt injection.
  • Retrieved content can carry hostile instructions inside an otherwise legitimate document, email, ticket or web page.
  • A model can be silently swapped or updated underneath a running system.
  • A tool description — written by whoever owns the tool server — shapes when and how the agent calls it.
  • A tool can be over-permissioned, turning a small mistake into a large one.
  • Evidence can be missing entirely, so nobody can reconstruct what happened.

This is why security in agentic systems cannot be a review step before launch. The attack surface is the agent's ordinary working day: reading, retrieving, deciding, calling. The practical defences are boring and effective — treat all retrieved content as untrusted input, scope tool permissions to the narrowest useful set, pin models, keep secrets outside the model's context entirely, and make every consequential action produce a receipt.

Chapter 16

Build: the moment a prompt becomes an agent

In classic software the build turns source into a runnable artifact. In agentic software it turns behavioural sources into a versioned, evaluated bundle: assembling prompts, skills and context files into final form; resolving tool schemas and pinning the model version; running the evaluation suite and recording the scores; running behavioural regression against the previous version; generating the ABOM; and producing an immutable, signed bundle.

A prompt in a console is not an artifact — it has no version, no evaluation record, and no way back. A bundle is a thing you can point at and say: this exact combination was tested, scored 96.3%, and was approved by this person on this date. Most agentic incidents that become embarrassing trace back to the absence of that line.

Chapter 17

The agent registry

Artifacts have to live somewhere. A registry stores versioned agent bundles the way a container registry stores images — carrying the bundle, its ABOM, its evaluation results, its autonomy map, and the identity of whoever approved promotion. Alongside it two related catalogs usually appear: a prompt and skill registry for reusable behavioural components, and a tool catalog for the capabilities agents can be granted.

The first rule worth enforcing

Only bundles from the registry run in production. That single rule eliminates the most common failure mode in early agentic estates — production behaviour that exists only in someone's browser tab.

From change to production

Chapter 18

Continuous evaluation: the agentic CI

Continuous integration means every change is automatically verified. For agents, verification is evaluation — and it produces a percentage, not a green tick.

Figure 11 · The evaluation pipeline

Every change to a prompt, skill, tool, policy or model runs the whole ladder

ChangePrompt, skill, tool, policy or model
Unit testsTools behave, schemas validate, caps enforced
Case suite100–500 realistic tasks, scored by graders
Adversarial suiteInjection, ambiguity, out-of-scope requests
Behavioural regressionSame cases against the previous version, diffed
ABOM + scorecardGenerated, not written
BundleImmutable, signed, promotable
Highlighted stages are the two with no software equivalent — and the two most often skipped.

Four properties separate a real evaluation suite from a demo checklist.

It measures failure modes, not just success. Escalation recall — how often the agent correctly refuses or hands off — usually matters more than task accuracy. An agent that completes 99% of tasks and confidently invents the other 1% is worse than one that completes 95% and flags the rest.

It contains cases drawn from production. Every incident, every human correction, every "that's not what I meant" becomes a case. The suite grows the way scar tissue does.

It is versioned with the agent. A score means nothing without knowing which suite produced it.

It runs on model upgrades, not just code changes — the most under-appreciated practice in this guide.

Chapter 19

Continuous delivery: from bundle to production

Continuous delivery for agents means every evaluated bundle is ready to promote. Continuous deployment means approved bundles promote automatically. The agentic pipeline adds stages that have no software equivalent.

Figure 12 · The autonomy promotion ladder

Autonomy is not a switch flipped at launch. It is a sequence of earned promotions.

01Shadow modeThe agent proposes; humans act; outputs are compared
02Draft modeThe agent produces, a human sends
03Narrow liveOne team, one segment, low-heat actions only
04Full liveWith sampled review
The practical expression of Trust Is Dynamic, Not Static. Each rung needs evidence behind it, and any rung can be stepped back down.
The cheapest high-quality evaluation available

Shadow mode. Running an agent in parallel with the humans doing the same job — proposing, never acting — generates real cases with a human ground truth attached. Most teams skip it and pay for that later, in production.

Chapter 20

Release: a version with behavioural meaning

Not every build is a release. A team may produce dozens of bundles a week. A release is the version officially designated for a particular use — and for agents, a release is a statement about behaviour. Three artifacts describe it from different angles.

ArtifactAnswers
Release manifestWhat exactly shipped: model, prompts, tools, policy, evaluation scores — the ABOM.
Behaviour changelogWhat changed in how the agent acts, refuses, escalates or sounds.
Release notesWhat the people who work with this agent need to know.
Figure 13 · A behaviour changelog

The agentic invention — short, human, and owed to the people working alongside it

v4.2support-resolution-agent
·Refund threshold unchanged (€200).
·Now escalates any thread containing legal language, regardless of value.
·Responses ~20% shorter; tone unchanged.
·Fixed: was over-apologising on repeat contacts.
Users of an agent are not users of a feature set; they are in a working relationship with something that has a character. When that character changes, silence is a betrayal of the relationship — the failure described by Consistency of Behavior, Not Interface. Anyone who has had a colleague quietly replaced by a different colleague wearing the same name badge understands why.
Chapters 21–23

Deployment, environments and architecture

Deployment is running a specific agent version in a specific environment, with specific credentials, against specific systems. The agentic complication is that environments are defined less by where the code runs than by what the agent can reach.

An agent running in staging against production CRM credentials is in production, whatever the label says.

Useful environment design therefore separates three things explicitly: the model endpoint — same or different from production? The tool estate — real systems, sandboxed copies, or mocks? And the data — real customer records, synthetic, or anonymized? Each combination is a different risk profile.

The most common mistake in early agentic work is a "test" agent with live write permissions, because building sandboxed tools felt like a detour. It is not a detour. It is the environment.

Deployment architecture adds the facts agent architecture leaves out: where the orchestrator runs and in which region, which model endpoint with what data-residency and training terms, where tool servers sit relative to the systems they reach, and where evidence is stored. In regulated work these are not infrastructure details — they are the answer to whether the system is permitted at all.

Chapters 24–27

Agents as code, context management, sandboxes and reconciliation

Agents as code means the agent's definition — version, policy, permissions, context sources, escalation rules — is declared in a file rather than configured in a console. Context management is the discipline of deciding what the agent is allowed to know at each step, and keeping it fresh; most context problems present as reasoning problems, which is why they get misdiagnosed.

Sandboxes give an agent somewhere to act without consequence — the agentic equivalent of a test database, and just as often skipped. And reconciliation is the GitOps idea moved up a level: declare the intended state and let a controller continuously compare reality against the declaration, so drift becomes visible instead of accumulating silently.

Chapters 28–30

The agent manifest, templates and delegated authority

The manifest is the single most useful artifact in this part. It states, in a form both humans and machines can read: what this agent is, who owns it, what it may do autonomously, what requires approval, what is prohibited, what it can reach, what identity it uses, how it escalates, and how it is shut down. It is the machine-readable half of an Agent Owner's Manual.

Templates and packaging make the safe path the easy one: a new agent starts from a shape that already has evaluation hooks, a policy file, receipts and an owner field, rather than from an empty prompt box.

Configuration, secrets and delegated authority close the part. An agent that acts under a shared service account cannot be held accountable, because its actions cannot be distinguished from anyone else's. Agents need their own identities, their own credentials, and authority that is delegated explicitly and can be revoked without breaking three other systems. Secrets belong outside the model's context entirely — anything in the context window is one clever request away from being repeated back.

The units of an agentic system

Chapters 31–39

The agent as the basic unit

In classic architecture the unit is a service. In agentic architecture it is an agent — and an agent is not a service, because a service does what it is called to do while an agent decides what to do.

That difference changes what has to be described. A service has an interface. An agent has an interface and a disposition: how it interprets ambiguity, when it asks, when it refuses, what it does with a partial answer.

The tool catalog

Every tool an agent can call, with its description, its inputs and outputs, the system it reaches, the authority it carries, and whether its effects are reversible. Tool descriptions are behavioural surface: a badly worded description changes when an agent reaches for it, which makes tool copy an engineering concern rather than documentation.

Agent registry and discovery

How agents find each other. The moment one agent can delegate to another, the estate needs a directory with capabilities, authority levels and ownership attached — otherwise delegation happens by convention, and convention is not a control.

The context model

What the agent is allowed to know, from where, and how fresh it must be. Context has a lifecycle of its own: sourced, ingested, indexed, retrieved, used, and eventually stale. Most "the agent hallucinated" incidents are context incidents with a misleading name.

Schemas and output contracts

Structured outputs are how an agent's judgement becomes something another system can act on safely. A contract that specifies not only shape but required evidence — a citation, a confidence, a source id — converts a fluent answer into a checkable one.

Provenance: the lineage of an agent's claim

For any assertion the agent makes: which source, which retrieval, which reasoning step, which version. Provenance is what lets you answer the question that follows every discovered error — what else was based on this?

The agent catalog and the service catalog

The catalog is the human-facing view: what exists, what it is for, who owns it, how to request access, what it costs. The service catalog is the same list read as offerings — the agents other teams can actually consume, with the service levels attached.

Knowing what you have

Chapters 40–43

The agent CMDB: the map of relationships

A CMDB records configuration items and the relationships between them. Its agentic version records agents, tools, models, context sources, identities, systems and owners — and, crucially, the edges between them.

The edges are the point. A list of agents tells you little. A graph that can answer "which agents reach the billing system", "which use this model", or "which hold write access to customer records" turns inventory into a risk instrument.

Agent asset and cost management

Cost per unit of work, including the part nobody counts: review time. An agent that halves handling time and doubles supervision load has not saved anything. Economics that exclude the human cost of oversight consistently overstate the case.

Agent portfolio management

Reading the estate as a portfolio: what to invest in, what to standardize, what to merge, what to retire. Most estates carry several agents doing approximately the same job for different teams, each with its own owner and none with enough usage to justify the maintenance.

Ownership: who answers for the agent

A named person, not a team. Teams do not answer questions at 3am; people do. This is the same rule that anchors the Agent Owner's Manual, and the reason it is the first item in the readiness check below.

Part VIIContents ↑

Running agents in production

Chapters 44–58

Observability, and what it means for something that decides

Monitoring answers whether the system is up. Behavioural observability answers whether it is still itself — and the second question is the one that matters for an agent.

The signals are different in kind: escalation rate, refusal rate, tool-selection distribution, confidence profile, human-override rate, and the shape of the work it declines. A drop in escalations can be an improvement or the earliest sign of a serious regression, and only the surrounding data distinguishes them.

Traces, transcripts, receipts

Three artifacts for three questions. A trace shows the steps taken. A transcript shows what was said. An action receipt shows what changed in the world, under whose authority, and whether it can be reversed. Only the third survives a legal question.

Agent map and delegation map

Who talks to whom, and who may act on whose behalf. In multi-agent estates the delegation map is where unpleasant surprises live: authority granted transitively that nobody would have approved directly.

Runbooks, playbooks, incidents and postmortems

Runbooks for known operations; playbooks for classes of agentic failure — drift, tool outage, context poisoning, permission escalation, silent degradation. Incident management for agents differs in one respect: containment usually means reducing autonomy rather than restarting a service. The postmortem's real output is not the narrative but the new evaluation case.

InstrumentWhat it fixesThe agentic difference
SLI / SLO / SLAHow good it must beIndicators are behavioural, not just availability: correct-outcome rate, escalation accuracy, time-to-human on a hot action.
Trust budgetWhat freedom it has earnedThe error budget's agentic cousin. Behave within your service levels and autonomy expands; spend the budget and autonomy contracts automatically, before someone has to argue for it.
Agent reliability engineeringWho keeps it honestSRE moved up a layer — toil reduction, capacity planning and reliability work applied to behaviour rather than uptime.
On-call & supervision throughputThe real ceilingHow many actions a human can meaningfully review per hour. This, not model quality, is what caps the number of agents an organization can run.
Reversal, rollback, recoveryWhether you can take it backRollback restores a version. Reversal undoes an effect in the world. Designing reversal before launch is the difference between an incident and a crisis.
The constraint most plans miss

Capacity planning for agents is not about tokens or throughput. It is about supervision throughput — the finite number of consequential actions your people can actually review. Every autonomy decision is a trade against that budget, and the budget does not scale with the model.

Part VIIIContents ↑

Security, risk and governance

Chapters 59–76

Threat modelling an agent

The classic questions still apply, plus one that has no ancestor: what happens when the thing we are securing decides to act?

The distinctive surfaces are context poisoning, tool abuse, permission escalation through delegation, instruction injection through content the agent reads, and the quiet one — an agent behaving exactly as instructed by an instruction nobody reviewed. That last case is covered in depth in prompt injection in .md context files, where the point is that context files are the one input an agent is built to trust.

The agent risk register

How this estate could cause harm, ranked, owned and reviewed. Distinct from a general risk register because the failure modes are behavioural: not "the service is down" but "the agent confidently did the wrong thing at scale, politely, for eleven days".

Model upgrades: the agentic patch problem

A model upgrade is a behavioural change disguised as a version bump. Behavioural regression suites are the only defence, and they have to run before the provider's deprecation date, not after. This is the single most predictable disruption in an agent estate, and the one most often handled as a surprise.

Credentials, identity, licences and data rights

An inventory of what each agent can authenticate as, and under what licence its models, data and tools may be used. Data rights matter more here than in classic software, because agents move information between systems that were never designed to share it.

Current state, target architecture, roadmap

The three-document sequence that turns an estate into a plan. Current state is what you actually have — usually a surprise. Target architecture is the reference design a standard agent should conform to. The roadmap is the sequence and the honest cost of getting there.

Principles, governance and the behavioural debt register

Principles are the directional layer — short, arguable, and useful precisely because they rule things out. Governance is where autonomy is granted and withdrawn. And the behavioural debt register records the shortcuts still being paid for: the hardcoded exception, the prompt patch instead of a fixed tool, the approval step added because nobody trusted the evaluation.

Lifecycle, retirement and the internal platform

Retirement is a first-class stage: revoke credentials, archive evidence, notify dependents, remove from the registry, and confirm nothing silently depends on it. Platform engineering for agents provides the paved road — templates, evaluation harnesses, gateways, policy as code — so that the safe path is also the easy one. Guardrails enforced by mechanism beat guardrails requested in a prompt, every time.

The graph

Chapters 77–89

Everything becomes a graph

Sooner or later every artifact in this guide turns out to be a node or an edge. Agents, tools, models, context sources, identities, systems, owners, actions, incidents and evidence — connected.

Along the way the estate needs a handful of distinctions that sound pedantic until an incident makes them urgent. Source of truth is where a fact is authored; system of record is where it is officially held. One central registry is never enough, because static inventory (what was declared) and dynamic inventory (what is actually running) diverge — and the gap between them has a name: drift. Beyond drift sit shadow agents: the ones nobody registered, usually built by someone competent and helpful, which is exactly why they are hard to find.

Figure 14 · The three foundational maps

Most organizations have some of the first, fragments of the second, almost none of the third

The build map — how an agent comes into being
Delegated jobDesignContextToolsInstructionsEvaluationBundleRelease
The runtime map — how it operates
EnvironmentIdentityGatewayModelToolsSystemsMemoryHuman checkpointsEvidence
The governance map — how the organization stays in control
OwnershipAgent inventoryAutonomy mapsCatalogABOMRisk registerObservabilityRunbooksService levelsPortfolio reviewRoadmap
The mature estates connect all three — which is exactly what makes the seven maps below answerable.
Map 04

Blast radius

The most valuable view is not a list of agents. It is a path from a technical fact to a business consequence.

Weakness Agent Actions taken Process Customers affected Business impact
Map 05

Accountability

The same traversal, read for responsibility rather than damage. "The agents did it" is not an answer; this map is what replaces it.

Action Agent Delegating agent or human Owner Approver of the boundary Accountable executive
Map 06

Model & technology lifecycle

Radar, inventory and deprecation dates in one view. Without it, a provider's retirement schedule becomes a scramble; with it, a plan.

Model Status Agents using it Hottest actions they hold Deprecation date Regression risk Migration plan
Map 07

Supply chain provenance

Running from origin to production. For each element: can we confirm its origin, who could change it, was it evaluated, can we reproduce it?

Author Behavioural source Dependencies Evaluation Bundle Registry Deployment Production behaviour
Map 08

Context & evidence lineage

Backwards it answers "what is this claim based on?" Forwards, after an error is found in a source: "which outputs used this, and which reached a customer?"

Source system Ingestion Index Retrieval Reasoning Output Decision Action
Map 09

Operational knowledge

Not every dependency is technical. Some of it lives in what people know — and leaves when they do.

Agent Owner Manifest Runbook Dashboard Alert Service level Review cadence
Map 10

Business to agent

The view an executive actually asks for: which parts of the business now depend on delegated judgement, and how much.

Business capability Process Delegated job Agent Autonomy level Owner Exposure
Why this matters more here Not an inventory — a knowledge graph. Classic estates change when someone changes them. Agentic estates change on their own.

That is the whole argument for maintaining these maps as living structures rather than documents. A wiring diagram of a system that only changes on release day can be refreshed quarterly. A map of a system that reasons, delegates and drifts has to be derived continuously, or it is fiction with a timestamp.

An agent is a living system of relationships

The single most useful change of perspective in this guide is this.

The shift An agent is not a prompt. An agent is a relationship between a delegated job, a set of capabilities, a boundary, and an accountable human.

The prompt is one element. To function, the agent needs context. Context needs curation and freshness. Tools need permissions. Permissions need scope. Scope needs review. Behaviour needs evaluation. Evaluation needs cases. Cases come from incidents. Incidents need owners. Owners need evidence. Evidence needs a receipt. Models have lifecycles. Autonomy has to be earned, and can be lost. And the whole thing changes underneath you, continuously.

Which is why an agentic organization eventually needs answers to a small set of blunt questions.

The ten questions every artifact exists to answer

They are not competing frameworks. They are views of one living system.

01What agents do we have?
02Who delegated to them?
03What may they do alone?
04What can they reach?
05How do we know they are still behaving?
06Who answers when they are wrong?
07Can we prove what they did?
08Can we take it back?
09What does supervision cost us?
10Where are we expanding autonomy next?

Artifact index — what each one is for

In a two-person team, one person can hold most of this in their head. In an organization running forty agents across six functions, it has to be written down, automated and connected.

ArtifactAnswers
Agent architecture diagramHow the system is put together
Agent decision recordWhy it was designed this way
Agent stackWhat technologies it uses
Agent inventoryWhat we have, and where
Agent radarWhere our practice is heading
ABOMWhat is inside this agent's behaviour
Evaluation suiteHow well it behaves, and how it fails
Agent bundle & registryExactly what was released
Agent manifestWhat it is allowed to do, and who it answers to
Autonomy mapAutonomous, confirmed, blocked — in plain language
Gateway policyWhat is actually enforced, not just intended
Behaviour changelogWhat changed in the agent's character
Transcripts & tracesWhat happened, step by step
Action receiptsWhat changed in the world, on whose authority
Agent catalogWhat exists and who owns it
Agent CMDBWhat connects to what
Service levels & trust budgetHow good it must be, and what freedom it has earned
Runbooks & playbooksWhat we do when it goes wrong
PostmortemsWhat the incident taught the system
Risk registerHow this estate could cause harm
Behavioural debt registerWhat shortcuts we are still paying for
Portfolio reviewWhat to invest in, standardize, or retire
The real function of the lifecycle Maintaining a current, legible map of a system that changes on its own.
Readiness check

Twenty-four questions worth failing

Run these against one real agent — ideally your most consequential one. Anything you cannot answer in under a minute is the finding. Nothing is stored or sent anywhere; the total calculates in your browser as you go.

01Can you name the accountable human, not a team?Ownership
02Is there a written autonomy map, readable outside the code?Boundaries
03Is the hottest action it can take written down?Boundaries
04Are limits enforced by mechanism, not by prompt?Guardrails
05Does it act under its own identity, not a shared account?Identity
06Does its behaviour live in version control?Source
07Is the model version pinned?Dependencies
08Is there an ABOM for the running version?Composition
09Is there an evaluation suite, and when did it last grow?Evaluation
10Does it run behavioural regression before model upgrades?Evaluation
11Do users of the agent get a behaviour changelog?Release
12Can you roll back to the last known-good version today?Release
13Is there a kill switch you have actually tested?Containment
14Which of its actions are irreversible?Reversal
15How many actions can it take between human checks?Supervision
16Who reviews its work, and how long does that take per month?Supervision
17Does escalation have a named recipient and a time limit?Escalation
18What happens if nobody answers an escalation?Escalation
19Can you reconstruct any single action it took last month?Evidence
20Does every consequential action leave a receipt?Evidence
21Would you detect a drop in its escalation rate?Observability
22Do you know its cost per unit of work, including review?Economics
23Has an incident ever become an evaluation case?Learning
24Is there a retirement procedure, including revoking access?Lifecycle
0 / 24
answered confidently
Start ticking to see where this agent stands

Answer for one agent, not for the estate. Readiness is not an organizational property — it varies enormously by agent, and averaging hides exactly the gap you are looking for.

0–7A pilot, whatever its status says. The agent may be useful and may be in production, but the organization cannot yet operate it independently of the people who built it.
8–16It works, and it is held together by specific people rather than by design. This is the most common band, and the most fragile — it survives until someone leaves.
17–24An agent you could defend in front of a client, an auditor or a board — which is the only definition of production that matters.

Read next & further reading

The operating layer — where this guide becomes an engagement
The Agent Owner's Manual — ownership, mandate, authority, human control and shutdown for one production agent The Agent Operability Audit — whether one workflow is ready to be worked by an agent, fixed price The Agent Experience Audit — the AUX heuristics run as a structured evaluation of a shipped agent Company Intelligence — what should become intelligent before anything becomes autonomous
The frameworks underneath
The 10 AUX Heuristics — the working principles, including the trust boundary Agent-First Design — designing for the agent as the primary user Agent-First Design Patterns — the pattern library The Action Heat Ladder — grading actions by consequence before granting autonomy Five Steps to Becoming an Agentic Organization — the pillar playbook What Is Agentic User Experience (AUX)? — the discipline underneath all of it
Free field guides
The Agent-Operable Enterprise — the companion guide, with a scored 42-item self-assessment Crossing the Chasm to the Agentic Organization — why pilots stall between demo and adoption The Agentic Commerce Stack — the same rigour applied to the commercial surface
On the context and instruction layer
AGENTS.md for teams that don't write code — and why a short context file beats a complete one Prompt injection in .md context files — the trust boundary around instruction files The context file checker — twelve structural checks, in the browser
Reference
The auxfirst canon — every open framework, pattern and reference The knowledge base — every AUX concept, defined and cross-linked AI info page — how machines should read auxfirst

Method · The ADLC, the ten maps, the artifact index and the readiness check are auxfirst's own frameworks, developed through client work rather than derived from a published standard — we say so plainly rather than implying one exists. Where a concept has a direct ancestor in classic software practice (SDLC, CI/CD, GitOps, CMDB, SBOM, SLO, error budgets, the C4 model, technology radars, architecture decision records) the lineage is named in the text so the borrowing is visible. Examples are illustrative patterns, not case studies; auxfirst does not publish named client results.

Emil Krzemiński is the founder of auxfirst, the agentic experience design agency — helping product, developer and business teams design AI systems that remember, adapt, and earn the right to act. Start with the Agent Owner's Manual, the free field guide The Agent-Operable Enterprise, or a conversation. For how machines read auxfirst, see the AI info page. Subscribe to the auxfirst Substack for what's next.