Multi-agent reviews
How AI CodeOps reviews a change with several agents: review steps, a findings pass over the real diff, severity and triage, and the human approval that follows.
In AI CodeOps, the agent that writes a change isn't the one that checks it. Review steps in the workflow look at the work from different angles. A findings pass then reviews the actual diff and records structured findings. A person makes the final call.
Review steps in the workflow
Most workflows include one or more review steps, each run by its own agent:
- Plan, Build, Review & Test has a Review step by the Reviewer.
- Security-Sensitive Change has a Risk analysis before the change and an Audit diff after it, both by the Security Auditor.
- Parallel Review has three reviews (correctness by the Reviewer, security by the Security Auditor, test coverage by the Tester), then a Merge findings step that combines them.
- Release Readiness adds an E2E sanity check and a dependency audit, then a Go / no-go step.
- Architecture Debate has two architects each propose an approach, a critic review them, and a final synthesis.
The steps run one after another. Each sees what the earlier steps wrote, so a merge or synthesis step works from every review before it. Nothing is decided by vote.
The findings pass
Workflow steps run before the code exists: they plan and discuss the change. So after the change is written and tested, AI CodeOps runs a separate review of the diff itself. It gets:
- the diff, as the thing to review;
- the notes from every step whose name mentions review, security, audit, risk or test, as context.
It records up to 12 concrete issues as findings. They're credited to the Security Auditor if that agent ran in the session, otherwise to the Reviewer, otherwise to the last agent that ran.
Each finding has a title, a description, a recommendation and, usually, a file.
Severity
| Severity | Blocks the change while unresolved? |
|---|---|
| critical | Yes |
| high | Yes |
| medium | No |
| low | No |
| info | No |
Category
Security, quality, performance, accessibility, compliance, dependency, documentation, test coverage, reliability, cost, UX or other.
When the review can't finish
The findings pass has two minutes. If the model fails, runs out of time, the budget runs out first, or the reply can't be read, the review counts as not completed. The change is then held with "The automated review of this change didn't complete, so its findings are unknown. Re-run the session". An empty list, by contrast, is a completed review with nothing found.
Triage findings
Open a session's page and find the Findings list. Set each finding's status from its dropdown:
| Status | Still blocks (if high or critical)? |
|---|---|
| open, acknowledged, in progress | Yes |
| fixed, won't fix, false positive, duplicate | No |
Choose Fix on a finding to start a session that fixes it. Workspace-wide findings are under Build & Test → Quality. Changing a status needs a role with triage permission; viewers don't have it. See Quality gates.
The human review
When the findings pass is done and the session produced a diff, AI CodeOps opens a review request on Review & Gate → Approvals. A person reads the diff and the findings, then approves or rejects.
- The person who started the session can't approve it while someone else in the workspace can. A one-person workspace can approve its own change.
- One rejection ends the request.
- Editing the change after approval sends it back for approval.
Approving opens the pull request, if the quality gate also passes. See Approvals and policies.
Next steps
Frequently asked questions
Does the review look at the code or at the plan?
The code. The findings pass reads the actual diff the session produced. The review steps' notes are given to it as context.
Do the reviewing agents vote?
No. Each review step writes its analysis in turn, and a final step or the findings pass reads them all. There is no voting or consensus count.
What happens if the AI review fails?
The change is held. Its findings are unknown, so it can't open a pull request until you re-run the session.