Summary
Cole Medin introduces his “Dark Factory” — a public experiment in building a fully autonomous software development system where an AI coding agent handles planning, implementation, pull requests, and production deployments without any human in the loop. The concept borrows from lights-out manufacturing (automated physical factories that have run without human presence since around 2001) and was applied to software by Dan Shapiro, whose blog post Medin references throughout.
The technical stack centers on Archon for orchestration, Claude Code for coding tasks, and Miniax M2.7 as the underlying model. The architecture is governed by three root-level documents loaded into context for every automated workflow: a mission.md defining what the application should do and what issues are in scope, a factory rules file specifying how changes are reviewed and merged, and a CLAUDE.md providing tech stack and architectural context. The underlying application being built is a RAG-powered YouTube content agent — publicly deployable, with GitHub issues automatically triaged and resolved by the dark factory pipeline.
Medin also walks through a maturity ladder for AI-assisted coding, from Level 0 “spicy autocomplete” through Level 4 fully autonomous operation, making this one of the more coherent frameworks available for reasoning about where AI coding assistance is headed. For developers interested in agentic systems, autonomous CI/CD, or pushing the boundaries of what tools like Claude Code can do unsupervised, this video is a detailed and honest look at both the promise and current limitations of the approach.
📺 Source: Cole Medin · Published April 16, 2026
🏷️ Format: Hands On Build
