Cole Medin tours the control dashboard for his AI software factory in this 16-minute development log.

Keep human decisions visible

  • His original system took a product requirements document, split it into issues, implemented them and merged changes automatically.
  • He reduced that autonomy after finding the approach inadequate beyond some proofs of concept.
  • The replacement surfaces blockers and owner decisions in one place, while agents handle the repeatable work.
  • Dependency mapping, parallel runs, staging and application validation still require an explicit process.

Inspect the factory as well as its output

  • A workflow map shows individual skills and decision points across two connected factories.
  • A decision log records earlier experiments so failed approaches are less likely to be repeated.
  • Version comparisons expose changes to workflow steps and prompts.
  • Run views show issue state, implementation, review and the complexity assessment that changes the workflow.
  • A scorecard tracks runs, merges and elapsed time for each factory version.

Keep tests and failures attached to versions

  • A test view lists feature and regression scenarios, their selected runs and their results.
  • Medin reports 21 workflow-related failures across hundreds of runs, plus two Claude Code rate-limit failures.
  • Those counts lack a precise denominator and independent quality assessment.
  • A newer rewind feature uses Neon database branches to revisit earlier factory versions and compare behavior.

Put execution history behind the dashboard

  • Archon runs the planning, implementation and validation workflows, including parallel worktree handling.
  • Agent turns, tool calls and decisions go into Neon Postgres for later investigation.
  • GitHub webhooks feed issue and pull-request events into the same database through Neon Functions.
  • Neon Realtime, described as a public beta, keeps SQL-backed dashboard views updated as events arrive.

Borrow the architecture, qualify the claims

  • The factory and Mission Control dashboard have separate public repositories.
  • The factory is also building its dashboard; Medin presents further self-improvement as an evolving goal.
  • “Hundreds of pull requests a week” is the creator’s throughput claim, without an audited production-quality result in this tour.
  • The video includes a paid Viktor segment and a Neon partnership disclosure.

“How can we have the minimal set of decisions for the human that gives the most amount of control possible.” — Cole Medin, 2:02