Case study · AI build system · 2026
A harness for getting the most out of AI and human work together.
How it works
- The spark: a human idea.
- Spec: the CTO agent, Fable.
- Spec review: another agent checks it against the real code.
- Build: Codex writes every line.
- Self-check: ruff, mypy --strict, pytest.
- Gate: full suite on a real database, zero skips.
- Review: another model tries to break it.
- Fails? Back to Codex, three rounds, then me.
- Merge: a second account approves.
- Shipped: and written down.
The crew is one part of a bigger system I run at home, with a brain for me, a brain for every project, and one assistant I talk to for all of it.
Faster, smarter and cheaper every day.
Every model is plug and play.
- Fable 5.1 writes the specs.
- GPT-5.6 Sol builds with Codex.
- Opus 5.5 runs the crew.
- GPT-6 Astra gives a second opinion, and the next model plugs right in.
Roadmap
shipped
- A full security stress test.
The crew tried to break itself on purpose and fixed everything it found.
- A live dashboard.
One place to see what every agent is working on.
next
- Ship a product start to finish on its own.
Designed, built, reviewed and shipped by the crew.
later
- A daily self-check.
An automatic audit that hunts for weak spots before they cause a problem.
- One coordinator across every project.
A single role keeping all the work in sync.
- Shared memory across projects.
Lessons from one build carried into the next.