Carry Intent Through Delivery
How product teams carry context, quality standards and ownership from a useful experiment into an operated product.

On this page
A prototype for a reporting tool looks ready. A reviewer can upload a sample file and receive a plausible summary. The team has not yet connected the current data source, tested missing figures or decided what happens when an integration fails.
The prototype has made the interaction concrete. Preparing it for customer work means connecting the right information, testing exceptions and deciding how people will recover when a dependency fails.
Keep the context with the work
The team needs a shared account of the user, the problem, evidence, desired outcome, constraints and quality standard. Record important alternatives and decisions that remain open. Keep it close to the work and update it as understanding changes.
A short narrative can explain why a decision matters; a prototype can expose an interaction; an eval set can define expected behaviour for representative tasks. A dependency map may make ownership visible. Choose the artefacts that help the team carry the decision into the work.
Label what a prototype demonstrates and what is simulated or missing. A test using sample data should not be mistaken for a system with permission to access live customer information.
Decide what ready means

Every release needs a clear owner, an intended result and a way to observe or recover from failure. The remaining evidence depends on the risk.
| Work | Questions before release |
|---|---|
| A reversible internal draft helper | Can users recognise and correct errors? Are data access and the limitations clear? |
| A customer-facing assistant | Does it meet the quality bar on representative tasks? Can users get help and can the team monitor failures? |
| An agent that changes records | Are permissions independently enforced? What requires approval? How are partial actions, duplicate attempts and recovery handled? |
| A high-consequence workflow | Have the relevant specialists assessed the evidence, controls, incident response and obligations? |
These examples are starting points, not a universal approval policy. Work with the people responsible for the domain and use the governance chapter to set an appropriate standard.
Plan review as part of delivery
A generated change still needs someone or something to check the aspects that matter. Automated tests can catch known errors; they do not prove that the requirement was right or that an untested failure cannot occur. Human review can supply judgement, but it requires access to the evidence, enough time and a clear standard.
Choose between automated checks, sampled review and individual approval based on the work. Estimate the review load before increasing the volume. If generated changes arrive faster than they can be checked, more output may only create a larger queue.
Release within a boundary
A release path might begin with offline tests, continue with a shadow or internal trial and then reach a bounded group of customers. Some work can safely follow a simpler path. Some should remain assisted indefinitely. Higher autonomy is a choice to justify, not a level every product must reach.
Before exposure, decide what will be observed, who will respond and how to stop or recover. For an agent, include partial completion, duplicate actions, permission failure, cost and escalation. Make the state visible to the people relying on it.
Keep ownership after release
The system will depend on context, integrations, permissions and people who handle exceptions. Name the outcome owner and the system owner. Agree who maintains source material and eval cases, handles incidents and decides whether the capability should continue.
Review what customers actually do, what fails and how much effort the new workflow creates. Feed that evidence back into discovery and the next delivery decision. Every Agent Needs an Owner provides a more detailed operating contract.
v3.1 · Updated Oct 2026