State, memory, and controlled loops

• Separate context, memory, and execution state.
• Design loops with limits, exit criteria, and recovery.
• Avoid agents that repeat actions or lose track of the thread.

Recommended minimum state

{
  "task_id": "inbox-2026-07-02-001",
  "intent": "crear_borrador_respuesta",
  "risk": "medium",
  "customer": "cliente@example.com",
  "attempts": 1,
  "approved": false,
  "next_action": "draft_email",
  "evidence": ["email original", "politica devoluciones"]
}

Healthy loops

• Attempt limit: never retry forever.
• Exit criterion: know when a response is sufficient.
• Escalation: request human help when there's no confidence.
• Idempotency: don't repeat external actions if the step has already been executed.