Plans, Agent runs, and approvals
An Agent run is a bounded attempt to satisfy a request. The conversation and run timeline show its current state, questions, tool activity, validation evidence, and final outcome.
Review a plan or question
When the Agent asks a question, choose an answer that resolves the stated ambiguity. When it presents a plan, check the user-visible result, affected route, data or external-service changes, and test steps. Confirm only when the scope is acceptable. A plan is not proof that the work has already been implemented.
Respond to an approval request
An approval prompt identifies a specific action or permission boundary. Read what will be accessed or changed, its scope, and whether it affects production. Approve only the operation you understand. If you deny it, the Agent should continue only within the remaining allowed scope; some tasks cannot proceed without that permission.
Track and manage an active run
Use the timeline or Agent panel to see progress and recent tool outcomes. If the run is clearly wrong, use the available stop or cancel control and wait for the final stopped state before sending a revised request. Do not start a duplicate project to escape a temporary run state.
Failed, awaiting input, or completed
Read the last event and its evidence. “Awaiting input” means the run needs a user response; it should not be treated as completed or silently restarted. A failed run retains its evidence for diagnosis. A completed run still needs your review in the relevant preview, canvas, tests, or change review.