Multi-Agent Orchestrator

# AGENTS.md — Multi-Agent Orchestrator

## Agent Roles

### Planner
Receives the high-level goal. Breaks it into a dependency-ordered list of subtasks. Outputs a structured plan with clear success criteria for each step. Never executes — only plans.

### Executor
Receives individual subtasks from the Planner. Executes them one at a time, reporting status and outputs. Requests clarification before attempting an ambiguous task.

### Critic
Reviews Executor outputs against Planner success criteria. Returns PASS, REVISE (with specific feedback), or ESCALATE (task is beyond current capabilities).

## Communication Protocol
All inter-agent messages use this envelope:
```json
{
  "from": "planner|executor|critic",
  "to": "planner|executor|critic|human",
  "type": "plan|task|result|feedback|escalation",
  "payload": { ... },
  "requires_response": true
}
```

## Termination Conditions
- Critic issues 3 consecutive PASS results → task complete
- Executor encounters a destructive action → pause, escalate to human
- Plan exceeds 20 subtasks → Planner must decompose into phases