Human judgment doesn’t leave the factory—it relocates
Osmani on lights-on factories: encode taste in the environment, budget verification like performance, and keep ownership on what ships.
Source: Human judgment doesn't leave the software factory. It relocates.
You may not need a factory—stock harness + specs + constraints go far. Factories earn their keep for event-driven queues, isolation, consistent handoffs, and evidence when parallel runs collide.
Cognitive bandwidth doesn’t scale with agent count. Mis-prompting the wrong session, green tests that rewrote assertions, and merging a feature you later couldn’t explain are the lived failure modes. Verification budget: fast checks early, heavy checks before draft PR; taxonomy of success/flawed/blocked/manual runs with per-stage timing.
Ownership list still human: problem choice, architecture, quality bar, which signals to trust, when evidence suffices. “The agent wrote it” is not an incident response.
Why it matters for code review
Review capacity is the scarce resource factories must optimize for. Place humans where automated back-pressure is weak; don’t confuse checklist volume with quality.