Teams & models
A team is the set of agents that run for a repo. Each role has a name, a CLI (harness), and a model. Roles coordinate over a peer message bus — there's no central orchestrator.
Roles
Built‑in role names carry meaning:
- architect — surveys and plans (optional; for larger work).
- dev — implements and commits.
- qa — reviews the diff before publish.
A role named anything else gets a custom prompt (use {repo.focus} to inject the
focus).
Presets
| Preset | Roles | When |
|---|---|---|
| Lean | dev only | fast, cheap, simple changes |
| Standard | dev + qa | most maintenance work |
| Full | architect + dev + qa | larger or riskier changes |
| Full mixed | arch + dev (opencode) · qa (claude) | mix models per role |
Fewer agents = faster, cheaper runs. More agents = more review and planning. For most repos, dev + qa is the sweet spot.
CLIs & models
AM is harness‑agnostic — choose per role:
| CLI | Notes |
|---|---|
claude |
API key or Claude subscription login; strong general coding |
opencode |
OpenCode‑compatible gateways/models |
pi |
provider‑prefixed models, e.g. opencode-go/deepseek-v4-flash |
tau |
lightweight harness |
Model format follows the harness (e.g. claude:claude-sonnet-4-6,
opencode-go/deepseek-v4-flash). Pick a capable model for feature‑ish work and a
cheaper one for mechanical tasks.
Always Test the exact model. Click Test next to a role to confirm the cli/model is reachable on a worker before relying on it — Verify readiness.