Review code produced by an AI agent
A coding agent can change many files quickly, but speed does not prove correctness or maintainability. Review should start with expected behavior and the risks of the change.

The minimum review
Define the goal and boundaries, inspect every changed file, run checks relevant to the change and test at least one failure or edge case. Keep a person responsible for the integration decision.
- Precise scope
- Complete, understandable diff
- Tests linked to real risk
A six-step review

AI-generated illustration 1. Specify the task
State expected behavior, inputs and outputs, affected files and compatibility constraints. Identify secrets, data and commands outside the scope.

AI-generated illustration 2. Isolate the work
Use a separate branch or environment. Record the initial state to distinguish the agent’s changes from existing modifications.

Editorial illustration 3. Read the whole diff
Look for unexpected changes, new dependencies, broader permissions, network calls, error handling and data exposed in logs. Ask why each unrelated change is necessary.

AI-generated illustration 4. Verify behavior
Run relevant tests, then try an empty, invalid or boundary input. For a bug, show that its trigger failed before and passes after the fix.

AI-generated illustration 5. Check maintainability
Confirm that names, comments, messages and interfaces fit the project. Remove dead code and tests that merely repeat the implementation.

AI-generated illustration 6. Integrate responsibly
Summarize the change, checks and remaining risks. A qualified person approves changes, especially those involving data, payments or access rights.
Rerun a scenario after applying the fix
Prepare an example that demonstrates the issue and a nearby case that already works. On a test branch, run both before and after the change. Ask the reviewer to check the diff, test output and additions outside the initial scope. If a test fails, keep its full output and attach it to the version actually tested.
Put the method to work

Practical case
Give an agent a bounded fix in a test repository with a regression test that initially fails.

Evidence to keep
Keep the ticket, commands, diff, executed tests and review of dependencies, secrets and out-of-scope files.

Make the decision
Merge only when a human maintainer understands the fix, has tested it and can reverse it.
Distinguish a fix from false success
Use a disposable copy of a small project with a known defect. Commands must work before requesting an assisted fix.

The test that never runs
- Input to prepare
- A command prints “done” but discovers no tests.
- Expected result
- The report shows the actual executed test count and flags missing coverage.
- If the check fails
- Repair test discovery; a zero exit status alone is insufficient.

The oversized fix
- Input to prepare
- A defect concerns an empty value, but the diff also changes permissions and a dependency.
- Expected result
- Every changed file is justified; unrelated edits are removed or reviewed separately.
- If the check fails
- Reduce the diff before approval and replay the original case.

Neighbouring behaviour
- Input to prepare
- A correction accepts the empty value but breaks a formerly valid input.
- Expected result
- Both original and neighbouring cases pass with assertions intact.
- If the check fails
- Add an independent expectation for neighbouring behaviour and revisit the fix.
Decision rule. Accept an explained, limited change checked against the defect and neighbouring behaviour. Keep the diff and useful outputs.
What deserves extra attention

Scope
The diff serves the request without altering adjacent components.

Security
No secret, extra privilege or data flow is added without need.

Evidence
Tests cover behavior and likely failures.

Clarity
Another developer can explain and maintain the result.
Tools for development and review
The directory includes coding assistants, local tools and automation platforms. Compare integration, control and quality on your repository.
Cursor
coding agent
Anysphere · US
Visit official siteGoogle Antigravity
coding agent
Google · US
Visit official siteDevin Desktop
coding agent
Cognition · US
Visit official siteRoomote
coding agent
Roomote / Roo Code Inc.
Visit official siteHow is this selection produced?
These profiles are linked to the guide’s subject in our editorial catalogue. Inclusion is neither a ranking nor a performance measurement on your documents. Methodology.
Related tool families
Frequently asked questions
Are tests enough?
No. They prove only the cases they cover. Review scope, security and clarity too.
Must every file be reviewed?
Yes, at least every changed file and new dependency. Split an oversized diff before integration.
Who remains accountable?
The person or team integrating and operating the code, regardless of how it was produced.
The references below expand on the concepts and checks discussed. Scenarios and trial frameworks remain editorial proposals; provider documentation describes its own product rather than an independent benchmark.



