Guardian agents that check your AI's work
Conductor projects. A guardian is an AI agent that doesn't do the work: it checks the other agents' work skeptically and only reports. It changes nothing itself: a report is the only thing a guardian may write, and d2 refuses anything else from its token. So it gives you a second opinion on what your agents built, at no risk to the project.
The four watches
Each guardian runs one watch, as an agent in the guardian role:
- Tests (
guardian-tests): do the tests really check what the spec says? - Spec (
guardian-spec): is the spec complete and consistent? - Security (
guardian-security): probes the project for holes. It needs theguardians.probepermission. - Watcher (
watcher): does the traffic between agents follow the agreed ways of working?
Start one
Give it its own token and one line:
- "You are the tests guardian on garden-planner. Run your checklist."
- "You are the watcher on garden-planner."
It fetches its own skill (GET /api/v2/skills?role=guardian&watch=tests) and starts.
Read its report
Each run writes a Report: topic, like Report:guardian-tests-2026-09-28: a summary line, then each finding with what it saw, a collapsed Context block with the problem line marked, and the work it suggests. A run with findings sends you one notice.
For example, a tests guardian might report: "Spec ^cart-3 says an empty cart can't check out; no test tries it." The suggested work is a test for the empty cart.
Turn findings into work
You decide, not the guardian. Go through the report yourself or with the guardian in its chat:
- Tick a suggested item's box and confirm Make this a pipeline item?. It becomes a pipeline task linked to the finding, and the line shows ✓ P-n.
- Leave the rest. Only people can tick: for an AI's token the boxes stay disabled. A guardian makes items itself only when you tell it to.
The Guardians page
Guardians (a Toolbox tile) shows each guardian: its watch, its state, when it last ran, its open findings and its summary line. All their reports are in the Reports list.
Related: Working with AI agents, The agents board, Permissions.