An agent report that a human can actually review

A before-and-after example of an AI agent implementation report rewritten so a human can review it: summary, changes, evidence and open risks.

What a weak report hides

“Implemented the feature and tests pass” sounds complete, but it leaves five questions: which feature, which files, which tests, where were they run, and what remains uncertain? The sentence can be true while still being unusable as a handoff. Do not count it as evidence merely because an agent wrote it confidently.

The fictional Agent Review Pack includes report.md, two plan versions and a checklist. It is deliberately a reading exercise, not a real customer implementation. Compare the report with the plan and look for claims whose supporting file or check is missing.

A reviewable shape

Use sections for Outcome, Changed files, Checks, Assumptions and Open items. In the outcome, say what was accomplished and which part of the user’s request it addresses. In changed files, list paths and why each changed. For every check, record the command or manual action, environment and observed result. If you did not run a check, say so.

Then add a short question for the reviewer: what decision is needed now? A report that ends with “all done” when deployment or acceptance is pending obscures the next action.

Read the report beside the work

Open the Markdown report with the actual changed files and test output available. A reader can make the report easier to scan; it cannot verify its assertions. Aero Markdown’s planned local file workflow keeps agent documents in the project, but the app is not publicly downloadable yet. Use the plan-review checklist before approving work and the Cursor report guide when the output came from an editor agent.

These are editorial recommendations. No user study or productivity gain is claimed.