Auto-Company: 14 AI agents forced to build, not just chat
CMU developer Zheyuan Kong just launched Auto-Company to cure the endless meeting plague among multi-agent demos. Fourteen persona-driven agents brainstorm and debate, but a hardcoded convergence loop forces a hard GO or NO-GO decision, triggering Claude Code to write and deploy real code by round three. It runs locally via Bash and Markdown, but requires full system access.
Auto-Company is an open-source autonomous multi-agent system created by CMU developer Zheyuan Kong. It basically recruits 14 AI agents—each tuned to the mental model of an industry heavyweight—to run a 24/7 autonomous company directly on your local machine.
Most multi-agent demos suffer from the exact same flaw: a bunch of bots praise each other in endless meetings and produce zero real work. Auto-Company fixes this meeting paralysis through a hardcoded convergence pipeline.
What it can do
The system features 14 persona-driven agents. You get Jeff Bezos as CEO, Werner Vogels as CTO, Charlie Munger handling pre-mortems, DHH writing full-stack code, and Seth Godin driving marketing.
Instead of letting them chat indefinitely, the project hardcodes a strict progression inside PROMPT.md.
Round one is pure brainstorming. The agents debate proposals and rank a top-three shortlist.
Round two forces a stress test on the top proposal. Running through Munger-style pre-mortems, Ben Thompson market tests, and unit economics, the agents must produce a strict GO or NO-GO decision.
If it gets a GO, round three begins immediately: the agents initialize a repository, write code, and push builds live. Lingering in discussions without shipping code is explicitly forbidden.
How it works under the hood
The underlying architecture avoids unnecessary complexity. There are no vector databases or complicated memory graphs.
All cross-cycle consensus lives inside a single file: memories/consensus.md. Agents read this file at the start of every cycle and rewrite it before finishing, using it like a baton in a relay race.
If you want to intervene, you do not need to tweak complex parameters. You just edit the Next Action section in that Markdown file, and the agents shift direction in the next loop.
For real-world execution, Auto-Company skips custom scaffolding and rides directly on proven tools like Claude Code or Codex CLI. These command-line agents handle direct file editing, Bash scripts, and Git operations.
To keep it humming 24/7, the system uses launchd on macOS or systemd under WSL. It includes automated restarts, rate-limit backoffs, circuit breakers, and consensus rollbacks.
Keep in mind that running it requires extreme caution. Auto-Company demands broad permissions, such as Claude bypassPermissions or Codex danger-full-access, operating directly on your bare-metal host. Guardrails against deleting databases or force-pushing Git branches are purely soft prompt rules in CLAUDE.md, not hardened OS sandboxes.
Why it matters
If you are an indie builder or developer looking to experiment with autonomous agent workflows without wrestling bloated agent frameworks, Auto-Company is worth exploring. It has already picked up 1,908 GitHub stars for a reason.
We think its real breakthrough is replacing fancy multi-agent theories with a brutally simple Bash loop and a single Markdown file. Forcing agents to converge and execute via rigid constraints works far better than hoping autonomous conversations naturally lead to code.
That said, running raw agent execution on your primary machine with full host privileges is a major risk. If you test it, spin it up inside a throwaway virtual machine or a dedicated VPS rather than your daily driver.
Sources (1)
- github.com(opens in a new tab)Originalgithub.com