How do I review an agent’s work before it’s marked done?
Add a review state between agent completion and done, with evidence, acceptance criteria and explicit approval for consequential work.
To review an agent's work before it is marked done, add a distinct awaiting-review state and prevent the executing agent from treating its own completion claim as acceptance. Require the result, checks and known limitations to be attached before a person or designated reviewer can approve it.
This separates two facts: the agent has stopped working, and the outcome has been accepted. They are not the same event.
Define the review gate
Before execution, write the observable criteria the reviewer will use. When the agent finishes, it moves the task to review with:
- the result or live URL;
- branch, commit and pull request where relevant;
- checks run and their results;
- screenshots or recordings for visual behavior;
- known limitations and checks not run;
- decisions that changed the original approach.
The reviewer compares this packet with the outcome and criteria. Approval moves the task to done; a finding returns it to in progress with a specific correction.
Scale review to risk
Low-risk formatting or generated changes may be accepted by automated checks. Authentication, billing, data deletion, public claims and production behavior deserve human review. Do not apply the same ceremony to every task, but make the rule explicit before the agent starts.
A Git-based workflow
Use protected branches and pull requests as the code-review gate. The agent opens the PR and reports checks, but cannot merge until the required review succeeds. Keep non-code acceptance criteria in the linked issue so the reviewer sees the whole outcome.
Where ecco fits
ecco keeps awaiting review separate from done and stores task evidence beside the original outcome. MCP-connected agents can submit work and request review; the person retains the acceptance decision and can send findings back without losing the task history.
Connect an agent to ecco. For a broader view of why the board sits outside any one assistant, read ecco versus Claude and ChatGPT.
Frequently asked questions
Can the same AI agent review its own work?
It can run a self-check, but that is not independent acceptance. Use another reviewer or a person for consequential work, especially when the executor interpreted ambiguous requirements.
What should happen when a reviewer finds a problem?
Record a specific finding against the task, return it to active work and require new evidence for the correction. Preserve the earlier evidence rather than restarting the history.
Can CI automatically mark a task done?
Only when the agreed definition of done is fully machine-verifiable and the risk justifies automatic acceptance. Passing CI often moves work to review rather than proving every product requirement.
What is the difference between finished and done?
Finished means the executor has produced a candidate result. Done means the required reviewer or policy has accepted evidence that the outcome was achieved.