Start with narrow roles

Planner: decides the plan and success criteria; does not edit.
Retriever: fetches context from files, RAG, or documentation.
Executor: applies scoped changes.
Verifier: runs tests, reviews diffs, and detects regressions.

TerminalCode

task:
  goal: "Añadir una validación sin romper API pública"
  max_total_steps: 16
  max_parallel_agents: 2
  memory_file: ".agent/state.md"

agents:
  planner:
    can_edit: false
    max_steps: 3
  executor:
    can_edit: true
    allowed_paths: ["src/", "tests/"]
    max_steps: 6
  verifier:
    can_edit: false
    commands: ["npm run lint", "npm test"]
    max_steps: 4

Minimal persistent memory

Memory should store decisions and state, not the entire chat. If you store noise, the agent will retrieve noise.

TerminalCode

# .agent/state.md
## Objetivo actual
Corregir validación de email sin cambiar contratos públicos.

## Decisiones
- No tocar base de datos.
- Mantener nombres de funciones exportadas.

## Evidencia
- npm run lint: pendiente
- npm test: pendiente

## Bloqueos
- Falta confirmar comportamiento con emails internacionales.

Circuit breakers for runaway loops

TerminalCode

loop_guards:
  repeated_tool_call:
    same_tool_same_args: 2
    action: stop_and_summarize
  no_state_change:
    steps_without_new_evidence: 3
    action: ask_human
  compute_budget:
    max_runtime_minutes: 20
    max_gpu_memory_percent: 90
    action: pause
  failed_command:
    same_error: 2
    action: change_strategy_or_stop

Official sources

Hermes Agent documentation
LangGraph Docs
LangGraph Persistence
Model Context Protocol