Cole Medin compares five AI software-factory designs in this 20-minute explainer, drawing on his own experiments and work with engineering teams.

Plan together, automate the middle, review the result

  • A software factory coordinates coding agents across planning, implementation, testing, review and deployment.
  • Medin’s usual workflow keeps people involved in upfront planning and final validation while agents handle a batch of issues.
  • Choose the review policy by how well tests capture the intended outcome and how much judgment the work requires.

1. Dark factory — broad autonomy, fragile intent

  • An informal request flows through planning, building, agent review and deployment without human approval.
  • Medin’s experiment produced changes that missed his intent even when the agents executed their instructions competently.
  • He reserves this approach mainly for prototypes whose results are easy to inspect and whose mistakes are tolerable.

2. Fleet factory — narrow work with clear checks

  • Apply a highly automated loop to bounded maintenance, refactors or migrations with explicit acceptance checks.
  • The workflow can resemble a dark factory; the difference is the tightly constrained work it accepts.
  • Strong test coverage can support greater autonomy, but passing the available tests still depends on choosing the right checks.

3. PR factory — a person reviews every change

  • Agents can implement issues in parallel, but each pull request returns to a human for approval.
  • Medin recommends this entry point for teams testing their build loop or handling work that needs close judgment.
  • Review remains a throughput limit, and batching implementation does little to reduce the number of human decisions.

4. Release factory — review a small integrated batch

  • Medin’s preferred balance starts with a jointly planned work order, typically containing 10–20 tickets depending on complexity.
  • Agents map dependencies, run independent work in parallel and use fresh sessions for planning, building and validation.
  • Separate agents check application behavior and review code; a lightweight manager watches for conflicts and blocked work.
  • Changes converge on a staging branch, where a person tests the application and reviews one combined merge request.
  • Problems become new issues for the same loop; production deployment follows human sign-off.

5. Self-improving factory — production creates the next work

  • Add a feedback loop that gathers production errors, traces and usage signals, then triages actionable findings into issues.
  • Keep the release factory’s build loop, staging environment and approval checkpoints underneath that feedback.
  • Medin describes this as an area he is exploring, with signal selection and triage tailored to each application.

Change the policy when the work changes

  • The same factory can require per-change approval for sensitive work and use stronger automation for a well-tested migration.
  • Medin’s five types are an operating framework, not a measured ranking of software quality or delivery speed.
  • The video includes a paid LangChain segment; its product claims are separate from the factory comparison.

“The builder never grades its own work.” — Cole Medin, 16:03