Multi-Agent e MCP — Orquestração Avançada de Agentes

Agente Principal (Scheduler):
  → Lê spec
  → Divide em subtarefas
  → Spawna subagentes

  Subagent 1: Lê src/api/users.ts
  Subagent 2: Lê src/db/schema.ts
  Subagent 3: Lê src/middleware/auth.ts
  Subagent 4: Lê specs/billing.md
         │
         ▼
  Subagentes retornam contexto ao principal
         │
         ▼
  Agente Principal: implementa com contexto completo

// .claude/settings.json
{
  "mcpServers": {
    "brave-search": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-brave-search"],
      "env": {
        "BRAVE_API_KEY": "YOUR_KEY"
      }
    },
    "sequentialthinking": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-sequentialthinking"]
    }
  }
}

## Fan-out
Spawn 3 subagents to search for:
1. All usages of function X
2. All imports of module Y
3. All tests related to feature Z

## Fan-in
After all 3 return, synthesize findings into a plan.
Then implement the plan in the main agent.

## Pipeline
1. Subagent A: Generate test cases from spec
2. Main: Review and approve test cases
3. Subagent B: Implement code to pass tests
4. Subagent C: Generate documentation from implementation

## Competitive
Spawn 2 subagents with different approaches:
- Subagent A: Implement with REST
- Subagent B: Implement with GraphQL
Compare results and choose the best.