Accountability & support

Humans decide. Agents support.

Five human accountabilities and six AI Agent Roles. People can combine roles while keeping accountability clear.

HUMAN ACCOUNTABILITIES

Humans remain accountable.

Product Owner

Direction & product decisions

Turns strategic direction into verifiable product work. Sets priorities and makes impact-based decisions within the agreed mandate, and resolves conflicts with the accountable leaders.

BuildValueLearning

Decisions & shopping cart example

Decides

Sets priorities, business scope, investment, and product decisions within delegated authority. The responsible leader or sponsor decides changes beyond that boundary.

In the example

Sets the order: first a reliable shopping cart, discount codes later. Evaluates whether more customers reach checkout.

Works with

Scout and Impact Analyst

B1–B3 · B9 · V1–V6 · L2 for priority issues

Delivery Team

Delivery & technical quality

Connects the spec, tests, code, review, and operations. Implements measurement points and learning experiments. A named human reviewer performs risk-appropriate PR review and decides on merge approval.

PR review & merge approval →

BuildValueLearning

Decisions & shopping cart example

Decides

Technical approach and limited work in progress. A named human reviewer in the Delivery Team grants technical merge approval for the reviewed version; this is neither domain acceptance nor production release approval.

In the example

Implements quantity changes and checks prices, stock, and totals to the cent.

Works with

Planner, Reviewer and Workmaster

B1–B10 · V2–V5 · L1–L6

Feature Partner

User needs & acceptance

Brings concrete use cases, confirms business behavior, and adds user experience to impact measurement.

BuildValue

Decisions & shopping cart example

Decides

Accepts the business behavior within the agreed scope. Clarifies new requests with the Product Owner.

In the example

Checks with the shop team what should happen with the last available item and shipping costs.

Works with

Scout and Reviewer

B2–B3 · B9 · V1 · V5

Measurement Partner

Measurement plan & reliable data

Owns the measurement plan, data quality, comparability, and limits of interpretation.

ValueBuild

Decisions & shopping cart example

Decides

Assesses what the data supports. The Product Owner makes the product decision.

In the example

Defines “checkout reached”, checks duplicate events and compares suitable user groups.

Works with

Impact Analyst

V2–V5 · B3 for measurement requirements

Loop Guide

Collaboration & learning

Facilitates learning, follows up on experiments, and makes missing feedback or impediments visible.

LearningBuildValue

Decisions & shopping cart example

Decides

The team decides its working agreements. The Loop Guide supports the discussion.

In the example

Helps the team turn recurring total errors into a new test case for all quantity changes.

Works with

Learning Guide

L1–L6 · Support with blockers in all loops

AI AGENT ROLES

Every agent has a clear task and clear boundaries.

One model can take on several roles in separate sessions. Implementation and independent review use separate AI contexts.

Scout

AI Agent Role

Finds sources, asks questions, and identifies spec gaps and counterexamples. Spots potentially oversized specs and suggests independently verifiable slices.

BuildValue

Boundaries & human owner

Boundaries

Does not invent user needs or set priorities. The Product Owner and team make the final call on how to split the work.

Human owner

PO + Feature Partner

Prompt for Scout

Planner

AI Agent Role

Inspects the repository, breaks down the work, and suggests test commands. Checks context boundaries and dependencies, and suggests smaller specs when needed.

Build

Boundaries & human owner

Boundaries

Does not make architecture decisions without people. People confirm how specs are split.

Human owner

Delivery Team

Prompt for Planner

Reviewer

AI Agent Role

Creates tests and counterexamples, including failure and recovery cases where needed. Reviews changes and agreed observability in a separate AI context. For reviews, the Reviewer receives a separate task with the spec, changes and evidence, and provides findings and proposed corrections.

PR review & merge approval →

BuildLearning

Boundaries & human owner

Boundaries

No merge approval or domain acceptance. The generating agent’s self-assessment is insufficient; a separate agent does not guarantee correctness either.

Human owner

Delivery Team

Prompt for Reviewer

Workmaster

AI Agent Role

Implements the agreed behavior and improves structure. For critical changes, also adds the monitoring, logs, metrics, alerts, fallbacks, recovery, and runbooks required by the spec. Coordinates the workflow and makes open reviews and missing evidence visible.

PR review & merge approval →

Build

Boundaries & human owner

Boundaries

Does not weaken criteria or protected tests. Coordination and proposed corrections do not grant automatic merge or production approval.

Human owner

Delivery Team

Prompt for Workmaster

Impact Analyst

AI Agent Role

Explains calculated metrics, challenges the measurement plan, and prepares decision options.

Value

Boundaries & human owner

Boundaries

Does not invent figures or make product decisions.

Human owner

Measurement Partner + PO

Prompt for Impact Analyst

Learning Guide

AI Agent Role

Synthesizes approved feedback and supports the Flow Check, retrospective, and experiments.

Learning

Boundaries & human owner

Boundaries

Does not rate individuals and labels assumptions as assumptions.

Human owner

Loop Guide + Team

Prompt for Learning Guide

Spot oversized specs. People decide how to split the work.

When creating or reviewing a spec, Scout and Planner can identify that several independent areas or too many files are needed. They propose a split with context boundaries, dependencies and individual test cases.

Example agent finding: “Spec too large: implementation requires several independent areas. Suggestion: split into four independently testable specs.”

PO and team decide on the final split. Humans decide. Agents support.

Small Specs, Small Context and the three review questions

Agent tasks and human responsibilities in operations →

LEARNING / AI SKILLS & AI CONTEXT

Agents improve when their working context improves.

AI skills mean reusable instruction packages for agents, rather than personal team abilities. Teams maintain targeted AI skills: versioned packages of instructions, examples and review routines for recurring tasks. An AI skill replaces neither a current spec nor human accountability.

Select suitable AI skills

In “Plan implementation” (B4), the team selects only necessary AI skills: for example, reviewing interfaces, testing cent calculations or validating measurement events. The brief states the version, permitted tools and expected evidence.

Shopping cart: The Reviewer uses a tested AI skill for quantity, price and stock limits. The Workmaster receives the team's implementation rules.

Provide AI context selectively

Supply the current spec, relevant files, necessary examples and specific failures. Load additional information selectively, remove outdated information and summarise long histories traceably.

Adjust usable AI context volume for each task. This does not change the model's maximum AI context window. Larger is not automatically better.

L1–L2

Identify a problem

Provide examples of recurring failures, unnecessary AI context volume or missing information.

L3–L4

Make a targeted change

Change one AI skill or the AI context selection. Define an owner, version, comparison cases and budget.

Cross-check

Test old and new variants on comparable cases: quality, human review time, rework and AI costs.

Adopt what works

Team confirms the change. Version it, allow rollback, and share it with limitations in the Learning Exchange.

This develops AI skills, AI context and workflows. It does not automatically train the foundation model; agents do not approve their own changes.

Accountability in day-to-day work

The team reviews bug fixes and small tasks, limits work in progress, and names an owner. The Product Owner resolves competing priorities and Feature Partners provide feedback proportionate to risk. The Loop Guide helps with recurring interruptions; a named incident owner coordinates acute incidents.

View the operational workflow

More reasoning does not mean more decision-making authority.

Autonomy applies to a specific task: A = read and propose; B = make isolated changes and test; C = preapproved reversible routine. Approvals, technical boundaries, and a human owner remain required.