Two applications, three loops
Product organisations meet agents twice. Once inside the team — research, synthesis, prototypes, PRDs, implementation, product operations — and once inside the product, where a customer hands a bounded task to an agent and expects a reliable outcome, clear controls and a way back when it fails. This bookshelf covers both, because the same purpose, authority, memory and evaluation books govern both.
Agents in the team
- How the organisation works: research, synthesis, design, specifications, implementation and product operations.
- The question is throughput and judgment — useful, source-linked work with a named reviewer.
Agents in the product
- What the customer experiences: task completion, useful automation, clear controls, reliable outcomes and recovery.
- The question is trust — correct completion, recoverable failure and cost per accepted outcome.
Three connected loops
Find what deserves to exist.
Product purpose, customer discovery, market insight, strategy, priorities and evidence.
Turn intent into working value.
Experiences, prototypes, PRDs, engineering, AI capabilities, releases and lifecycle ownership.
Learn from what actually happens.
Activation, retention, pricing, experiments, evaluation, reliability and organisational learning.
A purpose, a scope, an accountable owner, sources, an approved version, a review date, inputs, outputs, permitted actions, acceptance checks and escalation. A book can become a document, a schema, a playbook, a dataset, a workflow or an enforced configuration. Discovery, delivery and growth run in connected loops; the order of the shelves is a design logic, not a sequence to complete.
12 shelves, one question each
Each shelf answers one question and has a suggested owner. Every book names what the human owns, what the agent contributes, and the acceptance check that says the book is real. Shared foundation books stay visible in every discipline; the tags show the discipline and whether the book serves agents in the team, agentic product features, or both.
| Shelf | The question it answers | Owner | Books |
|---|---|---|---|
| 01 · Purpose & customer value | Why should this product exist, and which customer outcomes justify the work? | Product leader + founder | 6 |
| 02 · Strategy & portfolio | Where will we compete, what will we invest in, and what will we stop? | Product leadership + business owner | 8 |
| 03 · Authority & accountability | Who may decide or act, with which permissions and routes for intervention? | Product owner + relevant control owner | 6 |
| 04 · Team & operating model | How do people, product disciplines and agents coordinate around outcomes? | Product leadership + functional leaders | 6 |
| 05 · Discovery & customer evidence | What should we learn before deciding what deserves to be built? | Product trio + research lead | 8 |
| 06 · Design & product specification | What experience and behaviour are we committing to, including failure cases? | Product trio + relevant technical owner | 8 |
| 07 · Delivery & product lifecycle | How does an accepted idea become a maintained product in real use? | Engineering lead + product trio | 6 |
| 08 · Growth & monetisation | How do users reach value, keep receiving it, and support a viable business? | Product + growth lead + commercial owner | 6 |
| 09 · Product brains & shared memory | What knowledge is authoritative, retrievable and worth remembering? | A named owner for every context collection | 6 |
| 10 · Agent roles & orchestration | Which bounded work can agents own operationally, and how is it coordinated? | An accountable human owner for each agent | 8 |
| 11 · Systems & platform foundations | Where do agents run, retrieve context and interact with the product stack? | Platform / engineering owner + system owners | 6 |
| 12 · Evaluation, reliability & learning | How do we know the product works, and how do results improve the organisation? | Product + engineering + data / evaluation owners | 6 |
Purpose & customer value
Why should this product exist, and which customer outcomes justify the work? Suggested owner: Product leader + founder.
Purpose & mission
State the change the product should make for customers and the business. Give every initiative a reason beyond shipping a feature.
- Human owns
- Choose the mission and resolve competing commitments.
- Agent contributes
- Check proposals against the approved purpose.
- Accepted when
- A real initiative can be accepted or declined using the charter.
Customer outcomes
Define the progress customers seek, their current alternatives and the conditions under which the product is useful.
- Human owns
- Choose outcomes worth solving and validate them with customers.
- Agent contributes
- Organise evidence by desired outcome and unresolved assumption.
- Accepted when
- Each priority outcome links to observed customer evidence.
Value proposition
Explain who benefits, why the product is preferable in a specific situation and what evidence supports its promise.
- Human owns
- Approve the promise and the differentiated value.
- Agent contributes
- Compare alternatives and flag unsupported claims.
- Accepted when
- The promise states its intended audience, evidence and limits.
Business model & economics
Connect customer value, who pays, cost to serve, delivery capacity and the economic assumptions behind the product.
- Human owns
- Set business assumptions and acceptable trade-offs.
- Agent contributes
- Model cost and revenue scenarios with explicit assumptions.
- Accepted when
- The model includes downside scenarios and cost per useful outcome.
Product principles
Document the principles used to trade off simplicity, speed, user control, reliability and business goals.
- Human owns
- Choose principles and resolve conflicts between them.
- Agent contributes
- Review designs against concrete examples of each principle.
- Accepted when
- Principles explain decisions in at least three realistic trade-offs.
Users, buyers & stakeholders
Distinguish users, buyers, administrators, affected people and distribution partners. Their needs and authority may differ.
- Human owns
- Choose whose needs take priority in each decision.
- Agent contributes
- Identify gaps and connect research participants to their roles.
- Accepted when
- The map identifies excluded or underserved groups and decision owners.
Strategy & portfolio
Where will we compete, what will we invest in, and what will we stop? Suggested owner: Product leadership + business owner.
Strategic choices
Specify the target segment, problem, advantage, capabilities and explicit exclusions that give the team direction.
- Human owns
- Choose the position and resource commitments.
- Agent contributes
- Prepare options and evidence for the strategic trade-offs.
- Accepted when
- The strategy names choices, exclusions and conditions for revision.
Market & competitive dynamics
Examine customer alternatives, switching costs, distribution, market structure and changes that could weaken the product's advantage.
- Human owns
- Interpret the market and choose a response.
- Agent contributes
- Track sourced changes and distinguish observations from forecasts.
- Accepted when
- Every important market claim has a dated source and implication.
Opportunity portfolio
Compare opportunities using value, evidence, effort, dependencies and downside. Treat a score as decision support.
- Human owns
- Prioritise investments and make stop or defer decisions.
- Agent contributes
- Prepare comparable evidence and sensitivity to assumptions.
- Accepted when
- Each funded opportunity has an owner, uncertainty and next evidence gate.
Outcome metrics & guardrails
Define outcomes, leading signals and unwanted side effects with formulas, units, cohorts, time windows and authoritative sources.
- Human owns
- Approve the meanings, targets and decisions metrics support.
- Agent contributes
- Calculate from approved definitions and flag missing or incomparable data.
- Accepted when
- A second analyst can reproduce the result from the contract.
Outcome-based roadmap
Sequence problems, hypotheses and evidence gates while separating customer commitments from discovery options.
- Human owns
- Choose priorities and approve external commitments.
- Agent contributes
- Maintain dependencies and evidence behind proposed changes.
- Accepted when
- Every roadmap item has a problem, owner and decision checkpoint.
AI literacy & capability fit
Explain models, context, tools and agents in task terms. Compare an AI approach with rules, conventional software and human work.
- Human owns
- Decide where uncertainty and model behaviour are acceptable.
- Agent contributes
- Prepare representative task trials and compare quality, cost and latency.
- Accepted when
- The chosen approach beats a documented baseline on relevant cases.
Platform & distribution strategy
Choose how value reaches users through interfaces, integrations, marketplaces, partners or embedded workflows.
- Human owns
- Choose access routes and ecosystem dependencies.
- Agent contributes
- Compare channel constraints and prepare integration hypotheses.
- Accepted when
- Each route has a target user, adoption hypothesis and responsible owner.
Assumptions & pre-mortems
Record what must be true for a bet to succeed and imagine how it could fail before committing more resources.
- Human owns
- Decide which uncertainty deserves investigation first.
- Agent contributes
- Surface contradictions and update evidence behind assumptions.
- Accepted when
- The highest-risk assumption has a time-bounded validation plan.
Authority & accountability
Who may decide or act, with which permissions and routes for intervention? Suggested owner: Product owner + relevant control owner.
Decision rights
Assign accountable people for priorities, design choices, spending, customer promises, releases and incidents.
- Human owns
- Assign authority and resolve ownership gaps.
- Agent contributes
- Route decisions with evidence and the required approver.
- Accepted when
- Every consequential decision has an owner and escalation path.
Autonomy & human control
Set authority per action and context, including what an agent may read, propose, change or release and how a person can intervene.
- Human owns
- Approve scope, limits and changes to delegated authority.
- Agent contributes
- Act through granted permissions and pause at defined boundaries.
- Accepted when
- A denied or approval-required action cannot execute through the tool layer.
Data use & research permissions
Document allowed data collection, research participation, access, reuse and retention for each source and workflow.
- Human owns
- Approve applicable rules with responsible specialists.
- Agent contributes
- Apply documented rules and surface uncertain permissions.
- Accepted when
- Each source has an owner, permitted uses and a retention decision.
Security & service expectations
Specify expected service behaviour, access boundaries and failure consequences before choosing the architecture.
- Human owns
- Approve requirements and acceptable operational trade-offs.
- Agent contributes
- Identify requirements that proposed designs do not satisfy.
- Accepted when
- Requirements have measurable checks and accountable technical owners.
Promises & commitments
Control claims about product behaviour, delivery dates, AI capability and service levels, including budget and vendor commitments.
- Human owns
- Approve claims and commitments within assigned authority.
- Agent contributes
- Compare drafts and plans with approved evidence and constraints.
- Accepted when
- An external promise traces to approval and evidence or a stated condition.
Escalation & acceptance
Define what accepted work looks like, who resolves uncertainty and how blocked, failed or disputed work is handled.
- Human owns
- Accept material outcomes and resolve exceptions.
- Agent contributes
- Stop on unmet criteria and prepare a clear exception record.
- Accepted when
- A failed case reaches the right owner without being marked complete.
Team & operating model
How do people, product disciplines and agents coordinate around outcomes? Suggested owner: Product leadership + functional leaders.
Team topology & product trio
Map the partnership between product, design and engineering, with research, data and other specialists connected to the decisions they support.
- Human owns
- Choose team boundaries and assign outcome ownership.
- Agent contributes
- Expose dependencies and unclear handoffs.
- Accepted when
- Every product outcome has an accountable team and named partners.
Roles & mandates
Give every human and agent role a purpose, scope, inputs, outputs, decision rights and a named person accountable for its work.
- Human owns
- Assign roles and resolve conflicting responsibilities.
- Agent contributes
- Check task assignments against documented mandates.
- Accepted when
- No agent role is treated as the accountable approver of its own authority.
Planning & review cadences
Connect strategic planning, continuous discovery, delivery, growth experiments and results reviews in one operating calendar.
- Human owns
- Choose the cadence and make the decisions in each review.
- Agent contributes
- Prepare briefs, collect evidence and track open decisions.
- Accepted when
- Every recurring meeting has a purpose, inputs and a decision output.
Capacity & dependencies
Make people, technical constraints, agent workload, support load and cross-team dependencies visible when making commitments.
- Human owns
- Approve capacity allocation and negotiate trade-offs.
- Agent contributes
- Find bottlenecks and model the effects of changes.
- Accepted when
- Delivery commitments account for maintenance and dependency constraints.
Hiring & proof of capability
Translate interview and career themes into role scorecards, fair work samples and evidence of product judgment, collaboration and technical fluency.
- Human owns
- Make hiring decisions and assess human context.
- Agent contributes
- Prepare structured rubrics and organise evidence for human review.
- Accepted when
- The same job-related criteria are applied consistently with documented decisions.
Capability growth & feedback
Define capability levels and practical learning through coached product work, results reviews and feedback on decisions.
- Human owns
- Coach people and own development or performance judgments.
- Agent contributes
- Suggest practice exercises and organise examples from agreed work.
- Accepted when
- Development goals connect to observed work and a review conversation.
Discovery & customer evidence
What should we learn before deciding what deserves to be built? Suggested owner: Product trio + research lead.
Continuous discovery
Establish a recurring way to investigate opportunities with customers and connect learning to active product decisions.
- Human owns
- Choose learning priorities and maintain customer relationships.
- Agent contributes
- Prepare research plans and track unresolved questions.
- Accepted when
- Research leads to a specific decision or a revised assumption.
Interviews & field research
Plan recruitment, questions and observation around behaviour and context, with permission to use the resulting evidence.
- Human owns
- Run sensitive conversations and interpret context.
- Agent contributes
- Draft guides and structure permitted notes or transcripts.
- Accepted when
- Findings preserve source context and distinguish observation from interpretation.
Behavioural evidence
Use event data, support signals and permitted session observation to understand where users succeed or struggle.
- Human owns
- Choose questions and interpret the limits of the data.
- Agent contributes
- Identify patterns, data gaps and cases worth investigating.
- Accepted when
- Instrumentation quality is checked before a behavioural claim is accepted.
Synthesis & opportunity mapping
Connect observations to needs, opportunities and alternatives while retaining disagreement, uncertainty and provenance.
- Human owns
- Judge the strength and meaning of patterns.
- Agent contributes
- Group evidence, propose themes and link back to sources.
- Accepted when
- Every material theme is traceable and contradictory evidence remains visible.
Assumption tests & prototypes
Choose the cheapest credible test for a risky assumption, with a decision rule and the limits of what the test can establish.
- Human owns
- Choose the hypothesis and decide what the result means.
- Agent contributes
- Prepare test materials and compare outcomes with the rule.
- Accepted when
- The test can change a decision and has a recorded result.
Market & product case library
Study mechanisms behind other products and markets, recording what depends on context and what might transfer to your product.
- Human owns
- Decide which lessons are relevant to the current situation.
- Agent contributes
- Collect sourced examples and compare operating mechanisms.
- Accepted when
- A proposed lesson states its context and a local validation step.
Product-market fit evidence
Combine customer pull, repeated use, delivered value, retention and commercial evidence appropriate to the product's stage.
- Human owns
- Judge whether to persist, narrow, pivot or scale.
- Agent contributes
- Prepare segment-level evidence and flag conflicting signals.
- Accepted when
- A scale decision links to durable value evidence rather than a single score.
Feedback & support intelligence
Connect support conversations, customer requests and user feedback to product opportunities without equating request volume with priority.
- Human owns
- Resolve customer commitments and prioritisation judgments.
- Agent contributes
- Cluster feedback and route evidence to the right owner.
- Accepted when
- Customer issues retain identity, history and a visible disposition.
Design & product specification
What experience and behaviour are we committing to, including failure cases? Suggested owner: Product trio + relevant technical owner.
Journeys & information architecture
Design task flows, navigation, states and information structure around what users are trying to achieve.
- Human owns
- Choose the experience and key trade-offs.
- Agent contributes
- Generate alternatives and check state coverage.
- Accepted when
- The main journey and recovery routes can be walked through end to end.
Design systems & accessibility
Specify reusable interaction patterns, content standards and accessibility requirements across the product.
- Human owns
- Approve patterns and inclusive experience requirements.
- Agent contributes
- Check consistency and prepare component examples.
- Accepted when
- Representative flows pass agreed accessibility and usability checks.
Prototyping & usability
Build enough of an experience to test a specific question and record what remains simulated or unproven.
- Human owns
- Choose what to test and interpret observed behaviour.
- Agent contributes
- Prepare prototypes and organise usability findings.
- Accepted when
- A prototype result answers its research question without implying production readiness.
PRDs & acceptance criteria
Specify problem, intended users, scope, scenarios, constraints, instrumentation and acceptance criteria, including what is excluded.
- Human owns
- Own product intent and approve scope.
- Agent contributes
- Draft and reconcile requirements against evidence and constraints.
- Accepted when
- Engineering, design and QA can interpret the same examples consistently.
AI interaction & user control
Design initiation, progress, uncertainty, confirmation, editing, cancellation and recovery around the customer's task.
- Human owns
- Decide where users must choose, confirm or intervene.
- Agent contributes
- Prototype interactions and check coverage of ambiguous states.
- Accepted when
- A user can understand the task state and recover from an unacceptable result.
Context & output contracts
Define required context, authoritative instructions, input schemas, output schemas and evidence requirements for an AI task.
- Human owns
- Approve business meaning and required evidence.
- Agent contributes
- Assemble scoped context and return validated structured results.
- Accepted when
- Missing context or invalid output produces a defined failure state.
Model & capability selection
Compare candidate models and approaches using representative tasks, access constraints, quality, latency and total operating cost.
- Human owns
- Choose acceptable trade-offs and release criteria.
- Agent contributes
- Run controlled comparisons and report failure patterns.
- Accepted when
- The selected model has measured performance on the intended task mix.
Failure, fallback & recovery UX
Define what users experience when confidence, evidence, permissions or services are insufficient, including human support and recovery.
- Human owns
- Approve which actions stop and which fallback is acceptable.
- Agent contributes
- Detect defined failures and prepare the permitted recovery step.
- Accepted when
- Failure tests produce understandable outcomes without unsupported completion claims.
Delivery & product lifecycle
How does an accepted idea become a maintained product in real use? Suggested owner: Engineering lead + product trio.
Delivery workflow
Define how work moves from an accepted problem through design, implementation, review, release and follow-up.
- Human owns
- Choose the operating workflow and resolve priority conflicts.
- Agent contributes
- Prepare task state, dependencies and evidence for handoffs.
- Accepted when
- Work cannot be marked complete before its acceptance criteria are met.
Engineering & task decomposition
Translate approved requirements into bounded implementation tasks with dependencies, expected changes and verification.
- Human owns
- Approve architectural choices and review material changes.
- Agent contributes
- Prepare code or configuration within a scoped task.
- Accepted when
- Changes meet the requirements and meaningful checks cover remaining risks.
Version control & change history
Version code, prompts, skills, schemas, evaluations and approved context so a result can be traced to its released configuration.
- Human owns
- Approve changes and manage release ownership.
- Agent contributes
- Record versions and produce reviewable change summaries.
- Accepted when
- A production result can be linked to the relevant released components.
Release & progressive rollout
Specify who receives a change, readiness checks, monitoring, rollback and the evidence required to expand exposure.
- Human owns
- Authorise release scope and decide expansion or rollback.
- Agent contributes
- Prepare readiness evidence and execute approved rollout steps.
- Accepted when
- A limited rollout has observed results and a tested stop or rollback path.
Product Ops & handoffs
Maintain product records, decision inputs and handoff contracts between product, engineering, support and commercial teams.
- Human owns
- Assign system ownership and resolve service expectations.
- Agent contributes
- Maintain records and identify missing handoff inputs.
- Accepted when
- Receiving teams can use the handoff without reconstructing the decision.
Maintenance & sunset
Plan support, technical debt, compatibility, migration and retirement alongside new feature work.
- Human owns
- Approve lifecycle trade-offs and customer communications.
- Agent contributes
- Identify affected users and prepare migration or retirement checks.
- Accepted when
- A retired capability leaves users, records and dependencies in a known state.
Growth & monetisation
How do users reach value, keep receiving it, and support a viable business? Suggested owner: Product + growth lead + commercial owner.
Onboarding & activation
Define the first meaningful value milestone and remove barriers that prevent the intended users reaching it.
- Human owns
- Choose the value milestone and experience priorities.
- Agent contributes
- Analyse drop-offs and prepare improvement hypotheses.
- Accepted when
- Activation is defined by customer value with a consistent cohort and time window.
Engagement & retention
Distinguish useful recurring behaviour from empty activity, and understand retention by user need, cohort and product context.
- Human owns
- Interpret value delivery and prioritise interventions.
- Agent contributes
- Compare cohorts and surface changes requiring investigation.
- Accepted when
- Retention claims use appropriate cohorts and separate correlation from explanation.
Growth loops & distribution
Describe how delivered value creates the next adoption event, including referral, collaboration, content or ecosystem mechanisms.
- Human owns
- Choose the mechanism and acceptable customer experience.
- Agent contributes
- Model assumptions and instrument loop steps.
- Accepted when
- A loop has measured conversion, repeat behaviour and a clear failure point.
Pricing, packaging & AI costs
Connect customer value, packages, usage limits and cost to serve. For AI, account for variable inference, tool and review costs.
- Human owns
- Approve prices, packaging and commercial commitments.
- Agent contributes
- Model scenarios and flag margin or usage assumptions.
- Accepted when
- The model supports typical and heavy users without hiding cost assumptions.
Launch & product-led growth
Connect the product's value, activation, distribution, communication and support readiness for an intended segment.
- Human owns
- Choose the launch audience, narrative and readiness decision.
- Agent contributes
- Prepare launch assets and collect readiness evidence.
- Accepted when
- The intended customer can discover, try and reach value through a measured path.
Customer value & expansion
Use delivered value, adoption and unmet needs to inform account growth, renewals and product improvements.
- Human owns
- Own customer commitments and expansion decisions.
- Agent contributes
- Prepare value evidence and surface adoption or delivery gaps.
- Accepted when
- An expansion proposal explains the additional value and its evidence.
Product brains & shared memory
What knowledge is authoritative, retrievable and worth remembering? Suggested owner: A named owner for every context collection.
Product & strategy brain
Maintain approved purpose, product truth, strategy, principles, roadmap decisions and constraints in a versioned context index.
- Human owns
- Approve product truth and strategic changes.
- Agent contributes
- Retrieve the applicable context and flag contradictions.
- Accepted when
- Each important fact has an owner, source, version and review date.
Customer evidence vault
Store permitted interviews, observations and research artifacts with provenance, access rules, cohort context and uncertainty.
- Human owns
- Approve evidence use and interpret sensitive context.
- Agent contributes
- Retrieve relevant evidence without stripping its source or limitations.
- Accepted when
- A synthesis can be traced to original permitted research records.
Metrics & semantic definitions
Keep event meanings, metric contracts, entities, units, time windows and exclusions consistent across teams and agents.
- Human owns
- Approve business definitions and changes to meaning.
- Agent contributes
- Use the current definitions and explain comparability limits.
- Accepted when
- Two workflows answer the same metric question consistently.
Technical & design context
Index architecture decisions, design rules, interface contracts and known constraints needed for reliable product work.
- Human owns
- Approve architectural and design decisions.
- Agent contributes
- Retrieve relevant contracts and flag outdated references.
- Accepted when
- A task can find the current interface and design constraints it depends on.
Decision & experiment memory
Record what was decided, by whom, using which evidence and what would justify revisiting the decision.
- Human owns
- Accept learnings and authorise changes in direction.
- Agent contributes
- Surface relevant precedents and prepare proposed updates.
- Accepted when
- A repeated debate can recover the original evidence and decision conditions.
Context lifecycle & retrieval policy
Define approval, access, retrieval, freshness, conflict resolution and removal rules. Retrieved evidence does not grant authority to act.
- Human owns
- Approve authoritative context and retention decisions.
- Agent contributes
- Retrieve scoped evidence and report stale or conflicting material.
- Accepted when
- Tests exclude unauthorised sources and preserve provenance and current versions.
Agent roles & orchestration
Which bounded work can agents own operationally, and how is it coordinated? Suggested owner: An accountable human owner for each agent.
Skills & work contracts
Give every reusable agent skill a trigger, input contract, tools, output, permissions, acceptance criteria and stop conditions.
- Human owns
- Approve the mandate and allowed effects.
- Agent contributes
- Execute the defined skill and return evidence of its result.
- Accepted when
- A skill can be evaluated independently against representative tasks.
Research & synthesis agent
A proposed role for research preparation, permitted evidence analysis and source-linked synthesis supporting product decisions.
- Human owns
- Own research relationships and interpret material findings.
- Agent contributes
- Prepare research briefs and structured evidence summaries.
- Accepted when
- Important findings trace to sources and retain uncertainty or disagreement.
Product & design copilot
A proposed role for refining problem briefs, requirements, design alternatives and decision packs within approved product context.
- Human owns
- Own product judgment, experience choices and commitments.
- Agent contributes
- Draft artifacts and check consistency across requirements and evidence.
- Accepted when
- A reviewer can see assumptions, alternatives and the basis for each recommendation.
Implementation & QA agent
A proposed role for scoped implementation, test preparation and review assistance under repository and release controls.
- Human owns
- Approve architecture and release consequential changes.
- Agent contributes
- Prepare bounded changes and verification evidence.
- Accepted when
- The agent demonstrates behaviour against acceptance cases and reports failures.
Customer-facing product agent
A proposed product capability that pursues a customer's task with scoped tools, clear progress, confirmation and recoverable outcomes.
- Human owns
- Approve the customer experience and operating boundaries.
- Agent contributes
- Perform authorised task steps and report the actual outcome.
- Accepted when
- Representative customer tasks complete correctly or fail transparently with recovery.
Plans, loops & graphs
Choose a fixed workflow, bounded loop or branching graph according to task variability; define terminal states and iteration limits.
- Human owns
- Approve the control structure and consequential branches.
- Agent contributes
- Track progress, evidence and permitted transitions.
- Accepted when
- Repeated or ambiguous steps cannot create an unbounded execution loop.
Coordination & delegation
Use additional agents when separable work and evidence justify coordination cost. Define handoff contracts and independent acceptance.
- Human owns
- Own priorities, resource allocation and disputed results.
- Agent contributes
- Assign bounded subtasks and reconcile returned evidence.
- Accepted when
- Delegation preserves scope and cannot expand permissions or duplicate side effects.
Goals, checkpoints & budgets
Attach quality, time, spend and action limits to each run, with checkpoints, cancellation and resumable state.
- Human owns
- Approve budgets and decide when work should continue or stop.
- Agent contributes
- Track progress and pause at enforced limits.
- Accepted when
- An exhausted budget or cancellation causes a known, inspectable stopped state.
Systems & platform foundations
Where do agents run, retrieve context and interact with the product stack? Suggested owner: Platform / engineering owner + system owners.
Model & inference services
Register model services, versions, access constraints, performance expectations and supported fallback behaviours.
- Human owns
- Approve providers and operational trade-offs.
- Agent contributes
- Use the approved service configuration and record failures or versions.
- Accepted when
- A service change triggers compatibility and quality checks for affected tasks.
Tools, APIs, MCP & CLIs
Define callable capabilities, typed inputs, outputs, side effects and access scopes independently of a specific connector implementation.
- Human owns
- Approve capabilities and the identity allowed to call them.
- Agent contributes
- Invoke permitted interfaces and preserve action receipts.
- Accepted when
- A tool call can be validated, authorised and reconciled to its actual effect.
Data, identity & event contracts
Define entities, stable IDs, event semantics, access checks, duplicate prevention and reconciliation across systems.
- Human owns
- Approve authoritative sources and conflict rules.
- Agent contributes
- Map records and process events under those contracts.
- Accepted when
- Duplicate or delayed events do not cause duplicate business actions.
Storage, indexes & retrieval
Map operational records, files, search indexes and long-term context. Choose retrieval methods based on the question and evidence needs.
- Human owns
- Choose canonical stores and access or retention boundaries.
- Agent contributes
- Retrieve relevant authorised data with provenance.
- Accepted when
- Retrieval tests measure relevance, completeness and correct access boundaries.
Runtimes, queues & schedules
Specify execution environments, triggers, concurrency, retry behaviour, timeouts and state required to resume work safely.
- Human owns
- Approve service expectations and operational limits.
- Agent contributes
- Execute eligible work and checkpoint progress.
- Accepted when
- An interrupted run can resume without losing state or repeating completed effects.
Workspaces, harnesses & toolchain
Give teams controlled environments for research, prototyping, coding and agent work, with reusable context and clear environment boundaries.
- Human owns
- Choose tool families and approve environment access.
- Agent contributes
- Work within the assigned project, instructions and tool scopes.
- Accepted when
- Prototype, test and production environments have explicit access and promotion rules.
Evaluation, reliability & learning
How do we know the product works, and how do results improve the organisation? Suggested owner: Product + engineering + data / evaluation owners.
Evaluation cases & baselines
Build representative tasks covering normal, difficult and failure cases; compare results with a baseline and task-specific acceptance criteria.
- Human owns
- Define acceptable outcomes and review ambiguous cases.
- Agent contributes
- Run checks and report case-level evidence.
- Accepted when
- A change cannot pass by averaging away a material failure class.
Human review & judge calibration
Use explicit rubrics and human-labelled examples to assess automated judging. Measure disagreement and recheck after task or model changes.
- Human owns
- Own quality judgments and adjudicate disagreements.
- Agent contributes
- Apply calibrated rubrics and surface uncertainty.
- Accepted when
- Automated scores are checked against human judgments and known blind spots.
Experiments & causal evidence
Define the hypothesis, assignment unit, success metric, guardrails, exposure and analysis plan before interpreting an experiment.
- Human owns
- Approve the design and business decision rule.
- Agent contributes
- Check instrumentation and analyse results within the agreed plan.
- Accepted when
- The conclusion states uncertainty and what the design can establish.
Observability, quality & cost
Track task outcomes, traces, quality, latency, cost and human review load separately from activity volume.
- Human owns
- Choose thresholds and decide interventions.
- Agent contributes
- Link runs to outcome evidence and flag actionable deviations.
- Accepted when
- A failed or costly outcome can be traced to its relevant steps and versions.
Regression & release evidence
Check the combined effect of code, prompts, models, skills, data and tool changes before increasing customer or operational exposure.
- Human owns
- Approve release acceptance and any explicit exception.
- Agent contributes
- Assemble comparisons and report unresolved regressions.
- Accepted when
- The release decision records the tested configuration and remaining limitations.
Results reviews & incident learning
Connect product results, failed assumptions and operational incidents to proposed changes in strategy, design, workflows and agent instructions.
- Human owns
- Approve lessons and changes to the operating model.
- Agent contributes
- Prepare evidence, recurring patterns and follow-up tasks.
- Accepted when
- Each material lesson has an owner, an approved action and a follow-up check.
Give one agent a useful job. Give the organisation a way to learn.
Choose one product outcome or one team workflow. Prepare the books that make its context, ownership, permitted actions and acceptance clear. The suggested filenames are starting documents to create.
- Frame the work. Set the outcome and baseline. Choose a real task — turning interviews into a decision brief, helping a customer finish a bounded job, improving activation — and record how the work performs today.
- Prepare the minimum books. Agree context, authority and quality: an owner, approved inputs, a work contract, tools, limits and representative acceptance cases.
- Trial and review. Improve from accepted outcomes. Compare quality, customer value, time, cost and review effort. Resolve failures and missing context before widening scope or granting more autonomy.
Four project paths
Each is a reading order through the shelves, with the outcome it produces and the measure that says whether it worked.
Start one useful team agent
Outcome: One research workflow that produces a useful, source-linked brief with a named reviewer.
Ship a customer-facing agent
Outcome: A bounded customer task delivered through a clear experience with evaluation and a controlled rollout.
Improve product activation
Outcome: An evidence-backed change that helps intended users reach the first meaningful value milestone.
Design the product operating model
Outcome: A shared model of outcome ownership, decision rights, team interfaces and recurring review decisions.
The interactive bookshelf filters by discipline (Discovery, Design & delivery, Growth, AI & platform) and by application (agents in the team, agentic product features), loads a project path, and exports a build list or all eighty books as markdown.
How to set autonomy for a specific action
Assign autonomy to an action and a context, never to an agent as a whole. Review criteria, identity scopes, budgets and stop controls belong in the actual workflow and tool permissions — that is what the Authority & accountability shelf exists for.
| Mode | Working arrangement | Example |
|---|---|---|
| Assist | The agent recommends. A person decides and acts. | Preparing alternatives for a product strategy decision. |
| Co-pilot | People and agents develop the work together. | Refining a PRD or prototype with the product trio. |
| Delegate | The agent performs bounded work; consequential release requires approval. | Preparing and verifying an implementation change before release. |
| Autonomous within limits | The agent executes permitted actions under enforced controls; people handle exceptions. | Running an approved evaluation pack on a scheduled configuration. |
Where the course topics live
The two course maps that inspired this edition cover product management end to end. This is where each theme sits on the shelves — tool-heavy topics become capabilities and replaceable system roles rather than vendor choices.
| Source theme | Its place in the organisation |
|---|---|
| Product foundations and product sense | 01 Purpose & customer value; 02 Strategy & portfolio. |
| Strategy, roadmapping and competitive analysis | 02 Strategy & portfolio; 05 Discovery & customer evidence. |
| Customer discovery and customer intelligence | 05 Discovery; 09 Customer evidence and shared memory. |
| Product design, prototyping and PRDs | 06 Design & product specification; 07 Delivery. |
| Execution, delivery and product writing | 07 Delivery & lifecycle; 12 Results reviews. |
| Growth, user lifecycle, GTM and pricing | 08 Growth & monetisation. |
| Experiments and product analytics | 02 Metric contracts; 05 Behavioural evidence; 12 Experiments. |
| AI product strategy and building beyond the chatbot | 02 AI capability fit; 06 AI experience; 10 Customer-facing product agents. |
| Agent primitives, skills, memory, goals, loops and graphs | 09 Shared memory; 10 Agent roles & orchestration. |
| Research tools, prototypes, coding harnesses and team workspaces | 06 Prototyping; 10 Research and implementation agents; 11 Workspaces. |
| Context engineering, evaluations, LLM judges and observability | 06 Context contracts; 09 Context lifecycle; 12 Evaluation & reliability. |
| Company and market case studies | 02 Market dynamics; 05 A case library focused on transferable mechanisms. |
| Leadership and product organisation design | 03 Authority & accountability; 04 Team & operating model. |
| Careers, interviews and proof of work | 04 Hiring scorecards, work samples, capability development and coaching. |
How to use it — and what it is not
- Not a course. The maps informed topic coverage; the books are operating documents with owners, not lessons.
- Not a vendor assessment. Named tools appear only as examples in search metadata; this edition does not evaluate current features or recommend one.
- Not eighty documents on day one. One outcome or one workflow, the eight to ten books on its path, then the next.
- Scope extends to accessibility, permissions, operating economics, lifecycle ownership, recovery and organisational learning — because customer-facing agents fail there first.
The Product Agentic Organisation, v1.0, September 2026 — Revenue Puzzles working frameworks. The model applies to conventional products using agents internally and to products with agentic customer features. Sibling editions: the GTM edition and the whole-organisation edition. Book cards and the complete guide are generated from the same list.
The first session is one team workflow or one customer task, its baseline and the eight books on its path. Tell me what is going on and I will say which path fits.