Decide Architecture

Clasifica una tarea como WORKFLOW (deterministica) o AGENT (loop). Bias hacia workflow per Anthropic. Sugiere plantilla inicial.…

gonzalezpazmonica 42 updated 22d ago
Claude CodeGeneric
View source ↗
---
name: /decide-architecture
description: "Clasifica una tarea como WORKFLOW (deterministica) o AGENT (loop). Bias hacia workflow per Anthropic. Sugiere plantilla inicial. Mide accuracy contra corpus curado de 20 tareas."
developer_type: all
agent: task
context_cost: low
---

# /decide-architecture — Workflow vs Agent decision gate

> SPEC-158 — bias workflow-first per Anthropic effective-agents thesis.

## Intent

Cuando recibes una tarea nueva y no sabes si va por SDD workflow (deterministico, predecible, paralelizable) o agente autonomo (loop con tools, exploracion), invoca este comando antes de empezar. Te devuelve clasificacion + razones + plantilla sugerida.

## Sintaxis

```bash
/decide-architecture "task description here"
/decide-architecture --json "task ..."

Salida (texto)

DECISION: WORKFLOW|AGENT
workflow_score: N
agent_score: N
template: <ruta plantilla>
reasons:
  - <senal detectada con peso>
  - ...

Salida (JSON)

{
  "decision": "WORKFLOW",
  "workflow_score": 7,
  "agent_score": 0,
  "template": "docs/propuestas/_template-spec.md",
  "reasons": ["workflow:strong match 'deterministic' (+5)", ...]
}

Heuristica

Senal Peso Tipo
Verbos workflow: generate, compute, extract, format, parse... +1 WORKFLOW
Patrones workflow: deterministic, step 1, each file, for every... +5 WORKFLOW
Verbos agent debiles: understand, analyze, review, benchmark +1 AGENT
Verbos agent medios: explore, investigate, research, debug, decide... +2 AGENT
Patrones agent fuertes: loop until, iterate until, figure out, find the best... +5 AGENT

Bias: WORKFLOW empieza con +1 base. En empate gana WORKFLOW (Anthropic: "workflow first, agent only when necessary").

Plantillas sugeridas

  • WORKFLOW -> docs/propuestas/_template-spec.md (escribir SPEC SDD ejecutable)
  • AGENT -> .claude/agents/_template.md (definir agente con tools y loop)

Ejecucion

bash scripts/decide-architecture.sh "$@"

Validacion (AC)

scripts/decide-architecture-corpus-test.sh ejecuta 20 tareas curadas (10 workflow, 10 agent). AC: accuracy >=85%. Estado actual: 20/20 = 100%.

Cuando NO usar

  • La tarea ya tiene SPEC aprobada (entonces es claramente workflow).
  • El usuario indica explicitamente "delega a agent X" (skip clasificacion).
  • Tarea trivial (<5 minutos) — overhead innecesario.

Maintain Decide Architecture?

Let people know it's listed here — add the badge (live metrics, light/dark aware) or a plain link to your README or docs.

[Decide Architecture on getagentictools](https://getagentictools.com/loops/gonzalezpazmonica-decide-architecture-workflow-vs-agent-decision-gate?ref=badge)