Summary
Cole Medin explains why his experiment with an AI “dark factory” fell short of the dream: a fully autonomous coding system that takes a high-level epic or release document and ships to production with no human involvement. Drawing on Dan Shapiro’s five levels of AI coding autonomy, he argues that even with far better models, the bottleneck is communicating human intent, because no LLM can read minds.
The video then walks through the alternative he now uses, an AI software factory that keeps humans in control of planning while agents handle the core build-and-validate loop. Planning stays outside the factory, producing a PRD or architecture document that is split into bite-sized tickets. Each ticket then moves through a defined loop. The approach works for both greenfield proofs of concept and brownfield projects built on existing codebases, using Claude Code and Codex.
Medin also shows how a managed software factory can scale across large organizations, using Port’s context catalog and MCP server so agents know who owns which services and what standards apply. Developers will come away with practical lessons they can fold into their current agentic coding workflows, plus a view of where the industry appears to be converging.
📺 Source: Cole Medin · Published October 16, 2026
🏷️ Format: Workflow Case Study
