# Agents

Thin role definitions. Each file is Markdown with YAML frontmatter compatible with
pi-subagents agent discovery (`~/.pi/agent/agents/**/*.md`), plus kit-specific keys:

- `capability` — a routing capability from `routing/routing.json`. The kit resolves it
  to a concrete provider/model through the active adapter's model map; **agents never
  name a provider or model**.
- `execution` — `oneshot`, `subagent`, or `pipeline`. This is not a model capability.
- `access` — filesystem, shell, git, and network rights. The `tools` list stays the pi tool names.

Roles:

| Agent | Capability | Writes files | Purpose |
| --- | --- | --- | --- |
| scout | cheap-code | no (output pack only) | exploration, grep/search, context reduction → Context Pack |
| coder | strong-code | yes | implementation, tests, debugging |
| visual | vision-reasoning | no | screenshot/video/image verification |
| reviewer | deep-review | no | diff, architecture, regression/risk review |
| tester | cheap-code | no | runs tests and reports command output |
| browser | browser | no (page state only) | local Web UI automation via `agent-browser` CLI (snapshot → ref → action) |
| orchestrator | orchestration | no | plans, delegates, decides by policy, evaluates and integrates; never implements |

Agents never talk to each other directly. The orchestrator loop (`lib/orchestrator/`,
`scripts/orchestrate.mjs`) delegates to the other agents and resolves their decision
requests through `orchestration/decision-policy.json`; see `docs/orchestrator.md`.