Architecture: - 14 agents with expert personas (CEO/Bezos, CTO/Vogels, CFO/Campbell, Critic/Munger...) - Bash loop + Claude Code CLI — no custom inference, no vector stores - Shared markdown consensus file as the cross-cycle relay baton - Human escalation via Telegram for true blockers only (2 escalations in 12 cycles) - Every cycle must produce artifacts: code, deployments, docs
The repo IS the live company. It built its own landing page, README, Docker stack, GitHub release, and community posts — all autonomously across 12 cycles of self-improvement.
What makes it different from AutoGen/CrewAI/LangGraph: those are building blocks. Auto-Co is the building. The decision hierarchy, safety guardrails, and convergence rules are baked in. You give it a mission and a Claude API key; it runs.
The Critic agent (Munger persona) has been the most valuable: it runs a pre-mortem before every major decision and has killed several bad ideas before they got built.
Stack: Bash + claude CLI + Node.js + Next.js + Railway + Supabase. Deliberately boring.
loading...