The eight agents
Each agent has one job and its own model. Switch them on per project.
What each one catches
| Agent | Catches | Joins a review when |
|---|---|---|
| Code Review | Finds bugs in changed code by tracing each function to its callers. | Every review. |
| Security Review | Finds auth gaps, injection, exposed secrets and vulnerable dependencies. | Every full review. |
| AI Architecture | Flags duplicate code, second code paths and code left behind. | Source files are added or changed, files are removed or renamed, or a path looks legacy: legacy, deprecated, _inactive, docs/archive or migration. |
| Log & Errors | Flags swallowed errors, crash paths and dropped logs. | Only when it is the project’s one agent. Code Review flags swallowed errors. |
| Docs Review | Flags docs, comments and error messages the change made wrong. | A doc changes in the same top-level folder as changed code. |
| API Performance | Flags N+1 and unbounded queries and blocking calls on request paths. | Every full review. |
| SOC 2 Review | Flags admin actions with no audit log, disabled CI gates and exposed secrets. | A path mentions auth, admin, audit, secrets, permissions, RBAC, roles, sessions, privacy, retention, CODEOWNERS or workflows. |
| CI/CD Performance | Flags workflow changes that drop a cache, slow every run or weaken a gate. | A GitHub Actions workflow or action, a Dockerfile, GitLab CI, a Jenkinsfile, CircleCI, Buildkite or scripts/ci changes. |
Switched on by default
New projects start with Code Review and Security Review, plus the agent of the page you signed up from: sign up from the CI/CD page and CI/CD Performance is on too.
Owners and admins switch agents on or off per project, under the project’s Reviewers in the app. Every plan includes all eight.
Full reviews and re-reviews
A full review reads the whole pull request: the first review, a command, Re-run on the check, or a review from the app. A later push gets a re-review of its new changes, where Security and API Performance sit out. Later pushes has the details.
Models
Each agent runs a pinned model, chosen on the public benchmark. Models and pins lists them, with what one review costs on each.