Cole Medin builds a public AI "dark factory" that codes itself
Cole Medin is running a public experiment handing an entire codebase over to AI — using his open-source tool Archon and Claude Code to autonomously triage, implement, validate, and merge code with zero human-written lines.
Score breakdown
Developers and AI practitioners can study a fully public, end-to-end autonomous coding pipeline — including its governance layer and failure modes — to understand how to architect reliable agentic coding workflows with tools like Archon and Claude Code.
- 01Cole Medin is running a public dark factory experiment where AI agents write, review, test, and merge all code with no human-written lines.
- 02The target application is a RAG-powered agent that answers questions about his YouTube content.
- 03Archon, his open-source AI coding workflow engine, orchestrates triage, implement, validate, and fix workflows.
Cole Medin is publicly documenting a "dark factory" experiment — a software development setup where AI agents own the entire coding lifecycle of a codebase, from planning and implementation to pull request creation and merging, with no human-written code in between. The target project is a RAG-powered agent designed to answer questions about his YouTube content. Users (or Medin himself) file feature requests, and his open-source orchestration engine Archon runs the triage, implement, validate, and fix workflows autonomously using Claude Code and Minimax M2.7. The repository is live and all PRs are public, making the experiment fully observable.
The term "dark factory" originally described lights-out physical manufacturing facilities — robotic plants where no lighting is needed because no humans are present — a concept that dates back to around 2001.
The term "dark factory" originally described lights-out physical manufacturing facilities — robotic plants where no lighting is needed because no humans are present — a concept that dates back to around 2001. Dan Shapiro later applied the term to AI-driven codebases. Medin cites real precedents that validate the approach: StrongDM's Attractor project shipped 32K lines of Rust via AI, and Spotify's background coding agent accumulated over 1,500 merged agent PRs (documented in Spotify's engineering blog). Medin acknowledges the pattern is not yet fully reliable for production code, framing the effort explicitly as an experiment. His video covers the origins of the concept, levels of AI coding autonomy, his governance layer, the continuous factory loop, and the architecture he has built with Archon to make the approach as reliable as possible.
Key facts
- 01Cole Medin is running a public dark factory experiment where AI agents write, review, test, and merge all code with no human-written lines.
- 02The target application is a RAG-powered agent that answers questions about his YouTube content.
- 03Archon, his open-source AI coding workflow engine, orchestrates triage, implement, validate, and fix workflows.
- 04The experiment uses Claude Code and Minimax M2.7 as the underlying AI models.
- 05The repository and all pull requests are public and observable in real time.
- 06The 'dark factory' concept originated in lights-out physical manufacturing (~2001) and was applied to software by Dan Shapiro.
- 07Real-world precedents cited include StrongDM's Attractor (32K lines of Rust shipped by AI) and Spotify's background coding agent (1,500+ merged agent PRs).