I want to see what the agent built. This applies to solo work and to an AI Engineer’s work in a team. An AI message saying “done” does not replace a person viewing and checking the result. A small task can have a short demonstration; you still check that it created what was intended.
I stop the demonstration where I have feedback
I ask the agent to show me the path through the app with Playwright. In Codex I do this by voice: I watch the next steps and say, “Stop, I have feedback here. Add it to the feedback log,” or “Fix this now.”
You can also choose to record feedback and keep watching, or fix a small detail immediately and return to the same view. Voice is convenient for this kind of demonstration, but you can write too. Feedback goes to one place in the existing plan or task.
Show me a demo of the completed task [task or stage]. Use available Playwright or another prepared tool to show how it works. Walk through the user scenario one step at a time and stop for my feedback. When I ask to record feedback, add it to the list in the existing plan or task. When I ask for a correction now, correct that scope, check it according to risk, and refresh the same preview. Do not change the remaining scope. Use data allowed for the trial. Do not publish or deploy.
After the demo, return to the correction list
After the demo, the agent has a list of feedback to fix properly. For small changes, I do a quick review of the result. For larger ones, I ask for another demo focused on what we fixed. This lets me see how my feedback changed behavior.
After corrections, the agent runs needed checks. It checks changes to logic, data, and permissions immediately; a short demonstration does not free it from risk-appropriate tests. When you accept the whole demo, it tidies code, removes temporary workarounds, does the needed refactor and stage checks. If that changes accepted behavior, it returns to the demonstration.
Review the collected feedback list. Correct the agreed scope properly, check the effects of changes, and record the result for each note. Show me the corrected places again. For a larger change, walk through the entire changed scenario once more and say what you fixed. Do not mark my acceptance as complete before I see and accept the result.
For a stage without a screen, show a document, an answer with its sources, or the outcome of the action. The rule stays the same: a person checks what they received from AI.
Record acceptance of a view, the whole demo, and technical acceptance separately. Permission to publish is a separate decision. The end of a conversation alone does not mean acceptance.