Auto-Company runs 14 role-play agents as a 24/7 autonomous company on your own machine
aigclink · x · 2026-07-22
Auto-Company tries to stop agents from endless talking
The post argues that most “AI company” demos are exhausting because multiple agents just meet, praise each other, and never ship anything. Auto-Company takes a different approach: it runs a fully autonomous AI company on your own machine, with 14 agents modeled after well-known operators and builders.
Three notable ideas
- Forced convergence: the workflow is hard-coded in PROMPT.md. Round 1 is brainstorming and ranking the top 3 ideas. Round 2 adds validation through a pre-mortem, market test, and unit economics. From Round 3 onward, if the answer is GO, the system must create the repo, write code, and deploy. Endless discussion is explicitly banned.
- Memory as one Markdown file: instead of vector DBs or elaborate memory systems, each cycle reads and rewrites memories/consensus.md. Changing the Next Action there can redirect the next round immediately.
- Built on existing CLIs: execution is delegated to Claude Code or Codex CLI, using their native file, Bash, and Git abilities. A daemon layer on macOS/WSL handles restarts, backoff, rate limiting, and consensus rollback.
The catch is that it needs full permissions to run on the real host, with guardrails written as soft constraints rather than sandbox enforcement. The post’s main point is that the project solves agent empty-looping with a very practical, low-tech design rather than another “AI employee” pitch.
Related event: CMU Open-Sources Auto Company Framework with 14 AI Agents(3 posts)→
More from coding & agent
- Kolab Benchmark Uses Pro Screen Recordings to Show AI Agents Overestimate Collaboration Quality — minnesotanlp · 2026-10-27
- Dev argues Claude burns far more tokens than GPT models for the same coding work — haider1 · 2026-10-27
- Stop building agents that can't change models: stable contracts, external state, real evals — rseroter · 2026-10-27
- Cheap models can cost more in agent harnesses than frontier ones, researcher warns — vasuman · 2026-10-27
- Kanzo: A 5k-Line Rust Kanban Binary With MCP for Claude Code and Codex — allmudi · 2026-10-27
- 'A human approves everything' is a trap: approval fatigue and how to actually design safe agents — Efistoffeles · 2026-10-27