Review an agent's changes before moving on.

Keep the original task in view.

Open the task in Queues or its history. Read the instructions and the completion check. A status label alone does not tell you whether the requested behavior works.

In a synthetic task-app example, “reject whitespace-only task titles” needs two checks: blank input creates no task, and valid input still works.

Compare the change with its evidence.

Use Change Review for recorded project updates, Reports for findings, and Context for the session's outcome. Open the affected area in Project Graph or Files, then ask the agent to verify the current source and relevant tests.

• Does the change address the original request?
• Which checks actually ran, and what did they report?
• Are there failures, missing checks, or unrelated changes?
• Does the saved map need a refresh?

Graph comparisons show changes in published project metadata. Inspect source changes with your agent's existing repository access.

Review the completed task [task] in Brain Scanner project [name and ID]. Compare the instructions with the recorded outcome and current source. Report the checks actually run, remaining failures, and missing evidence. Keep source files unchanged during this review.

Recover after an agent breaks a feature.

Start with one reproducible failure. In our task-app example, valid titles stopped saving after a validation change. Compare the last submitted session's goal, outcome and checks with the current behavior. A reported passing check may not cover this input.

1. Ask the agent to reproduce the failure in its local checkout without editing yet.
2. Inspect the relevant diff when supplied, or have the agent inspect its local Git diff. Identify the smallest change that explains the failure.
3. Choose a targeted fix or explicitly approve reverting that change. Preserve unrelated work.
4. Check both the regression and the original requirement: valid titles save, blank titles do not.

In project [name and ID], valid task titles stopped saving after [session/change]. Read the recorded decisions and checks, then reproduce both valid and blank inputs locally. Explain the cause and propose the smallest repair before editing. After the approved repair, report the checks you actually ran and any remaining failure.

Use the selected session's decision trail to carry the evidence into your next session. It records submitted context, not a replay of everything the agent did.

Separate finished work from the next task.

If the change meets its completion check, keep the result in history. If something remains, edit an active task or queue a specific follow-up with its own check.

For a task-title fix, the validation logic can be checked separately from the form's error message and keyboard behavior. The latter becomes the next focused task.

Ask your connected agent to pick up that task. The queue organizes work; execution uses the agent's existing access and permissions.

Leave a useful outcome.

Ask the agent to record the changed area, actual check results, and unfinished work before ending the session. On your next visit, begin with that outcome and verify it against the current project.

Resume with the same or a different agent · Trace a change through callers and tests

Review your project