Autopilot V3 with AI: how we are automating Berraquero.com without handing over production
Updated: 2026-05-26
We are building an AI autopilot for Berraquero.com, but not in the reckless way people usually sell online. This is not about throwing an agent at production and hoping for the best. It is about turning repetitive work into controlled jobs with checks and human decisions where they matter.
What we are building
Autopilot V3 is not a bot that pastes commands and brags about speed. It is an operational layer for turning repetitive work into jobs with intent, declared risk, acceptance criteria, verifiable output and reviewable evidence.
The difference is not that an AI can run commands. That already exists. The difference is whether the system knows when it can act, when it must stop and when Toni must decide.
What we learned before V3
V2 proved that the flow could work: ChatGPT, browser extension, bridge, AutoHotkey, PuTTY and returned output. It also showed its limits: window focus, long outputs, timeouts, repeated blocks and too much logic inside one script.
The architecture that makes sense
The serious version separates responsibilities: Jarvis coordinates, Codex reviews code, GitHub versions, PostgreSQL stores state, n8n runs workflows and Autopilot validates server tasks. Toni decides the sensitive parts.
Real risks
The real risk is not that AI is dumb. The risk is that it sounds confident enough for you to give it more permission than it should have. A bad recommendation wastes time; a bad production action can break, duplicate, delete or publish the wrong thing.
Common mistakes
- Confusing logs with real validation.
- Mixing build, restart and public checks without need.
- Using n8n as a hidden brain.
- Giving permissions without acceptance criteria.
- Not storing reviewable evidence.
Quick tips
- Start with read-only tasks.
- Separate code, server, workflows and human decisions.
- Define objective, risk, timeout and expected proof.
- Do not accept partial success in critical jobs.
- If it touches production, it must be easy to explain clearly.
You can support the project or share this article in one click. At least this block does something useful.
Content reviewed using experience, authority and trustworthiness criteria (E-E-A-T).
FAQ
| Capa | Función | Por qué importa |
|---|---|---|
| Jarvis Core | Coordina intención y agentes | Evita que cada herramienta mande por su cuenta |
| Autopilot | Ejecuta y valida trabajos de servidor | No convierte un OK en verdad absoluta |
| Codex/GitHub | Código, revisión y trazabilidad | No se toca producción como una libreta |
| PostgreSQL | Estado operativo | Permite saber qué está pendiente, cerrado o bloqueado |
| n8n | Workflows persistidos | Ejecuta, pero no gobierna |
¿Autopilot V3 sustituye a Toni?
No. Quita carga repetitiva, no criterio humano.
¿Es un agente IA normal?
No exactamente. Es una capa operativa con carriles, evidencias y límites.
¿Puede tocar producción?
Solo bajo condiciones controladas y con validaciones reales.
Experiencia y criterio editorial
Este artículo sale del trabajo real de construcción de Berraquero.com. No es una demo aislada ni una promesa de laboratorio.