​
Definition

Review judges whether an output satisfies the request. A reviewer inspects the Issue, output, Agent Run evidence, and comments, then records a structured decision with a required comment. Accepted work may complete. Requested changes return to the assignee, while a blocked result keeps the missing input visible. Approval is different from Review. An Approval authorizes a governed action, such as spending or a sensitive change. A Review can accept the quality of an output without authorizing the action that uses it, and an Approval can permit an action without judging the quality of its output. Feedback names what was good, wrong, missing, or worth repeating. Learning is the deliberate promotion of that evidence into future context, a Skill, instructions, a workflow, a decision, or an evaluation case.

​
One illustrative case

A writing Agent submits a launch announcement. The reviewer accepts the setup instructions but requests a clearer limitation paragraph. After the revision is accepted, the team notices that the same limitation check appears in every release review. It drafts a change to the release Skill, cites the reviewed Issues, and evaluates the update before enabling it for later work.

​
When each step is useful

Use Review when completion depends on judgment by someone other than the assignee. Use feedback whenever the next actor or a later run should understand the reason for the decision. Promote a lesson only when the pattern is stable enough to improve future work beyond the current Issue. Reviewer routing and the final reviewer decision are separate events. Moving an Issue into a reviewable state gives the reviewer the next action; the work is not accepted until a structured decision is recorded.

​
Operating boundaries

An assignee Agent cannot bypass a named reviewer by marking the Issue done. Free-form accept or reject text is useful evidence, but it is not the structured review decision. Review follow-up is bounded and remains owned by the reviewer. The highest-risk boundary is silent learning. Feedback never rewrites durable Agent behavior automatically. A learning change needs linked evidence, a clear target and scope, an appropriate approval path, and a way to evaluate or roll back the change. Follow Review Agent work for the procedure. Use Approvals, budgets, and activity for governed-action controls.