Agent roles and responsibilities

Zapat uses a multi-agent team model where different agents are responsible for different parts of the pipeline. Each agent uses a model tuned for its task.

Triage Agent

Triggered by: "zapat:triage" label or @zapat mention on an issue. Responsibility: reads the issue title, body, and labels; classifies the issue type (bug, feature, refactor, research); estimates complexity (solo = one agent, duo = two agents, full = full team); and writes a structured plan that other agents follow. The triage summary is posted as a comment on the GitHub issue.

Implementation Agent

Triggered by: "zapat:implement" label (or automatically after triage). Uses the most capable model in Zapat's stack. Responsibility: reads the triage plan and relevant source files, writes the code changes, commits to a new branch, and opens a pull request. The PR includes a structured description explaining what changed and why.

Review Agent

Triggered by: "zapat-review" label on a PR, or automatically after implementation. Responsibility: reviews the diff for correctness, code quality, naming conventions, and potential issues. Leaves inline comments on specific lines where changes are needed, and a summary review. If the review results in "changes requested", Zapat dispatches a rework job.

Rework Agent

Triggered by: "zapat-rework" label on a PR, or automatically when a reviewer submits "changes requested". Responsibility: reads the review comments, implements the requested changes, and pushes a new commit. After rework succeeds, the pipeline always returns to the Test stage next — it never goes directly to review. This sequential guarantee (rework → test → review) prevents concurrent test and review jobs running against the same PR. The pipeline supports up to 3 rework cycles by default (configurable up to 10).

CI Fix Agent

Triggered by: CI failure on a Zapat PR (when "ciAutofixEnabled" is true). Responsibility: reads the CI failure logs, identifies the root cause, applies a targeted fix, and pushes a new commit. The pipeline supports up to 2 CI fix attempts by default (configurable up to 5).

Research Agent

Triggered by: "zapat:research" label. Responsibility: investigates a technical question or codebase area and posts a structured summary as a GitHub comment. Useful for answering questions like "how does the authentication system work?" before implementation.

Test Writer Agent

Triggered by: "zapat:write-tests" label. Responsibility: reads the code related to the issue and writes comprehensive tests — unit tests, integration tests, or both depending on context. Opens a PR with the new test files.