Cole Medin explains his revised AI software factory in this 19-minute talk on his channel. Agents handle implementation; people own the release plan and final acceptance.

Why he changed the design

  • His original “dark factory” took a broad application brief and aimed to ship directly to production.
  • Better coding models still make assumptions about what customers want and how a codebase should evolve.
  • He keeps people involved in product decisions and maintenance while agents work through a bounded release.

Plan one release

  • Define a proof of concept for a new app, or a small bundle of changes for an existing one.
  • Medin targets roughly 10–20 tickets, with explicit requirements and exclusions.
  • Create holdout scenarios: acceptance tests hidden from the implementing agent to discourage coding only for visible tests.
  • Map dependencies, then label tickets so independent work can run in parallel.
  • Choose human approval points by risk: regulated changes may need ticket and pull-request review.

Run the build loop

  • Each run claims one issue and produces a pull request, a proposed code change.
  • Route simple tasks through a fast workflow; reserve deeper iteration for complex work.
  • Validate behavior by using the app: click buttons, enter text, and check the outcome.
  • Review code separately for simplicity, maintainability, and sound implementation.
  • Send failures back for repair, with retry limits; merge into staging only after review and automated checks pass.

Supervise the factory

  • A manager agent checks the whole queue every 2–3 hours for stuck work and changed dependencies.
  • It can realign tasks, rescue a run, or ask a person to intervene.
  • Protect mission.md, engineering.md, and global rules from factory edits so agents cannot rewrite their own boundaries.
  • Use different models for different steps. Medin reports hundreds of weekly runs within rate limits, without a measured total-cost breakdown.

Test the release in staging

  • People test the assembled release in staging, a pre-production environment, to check whether it matches their intent.
  • Turn feedback into new issues; the same factory implements the corrections.
  • Repeat until the release is ready, then approve production and plan the next batch.

“We don’t want the agent to assume our intent even if it is a very capable model.” — Cole Medin, 6:24

Caveat: this is a sponsored workflow overview with self-reported reliability, not a comparative evaluation.