The meta-loop
The meta-loop is the feedback the engine generates while building Black Box itself, fed back into the engine's playbooks and prompts. Every shipped feature emits lessons; every brittle playbook surfaces a fix; every owner override teaches the CEO when to ask vs. act.
In one breath
- The engine ships its own engine; build lessons feed back as prompt and playbook edits.
- Owner-override patterns and evaluator-rejection clusters are first-class signals.
- Closed weekly: review, fold in, ship, tag the resolved lesson.
Why we run the engine on ourselves
Every product team should eat its own dog food, but for an autonomous agent platform that is the whole point. We use the engine to ship the engine. When a specialist gets stuck building a feature, that stuck-state is a fix-it ticket against the engine itself.
What the loop captures
Three streams. Build lessons: bottlenecks, recurring frustrations, prompt edits that fixed common failures. Owner-override patterns: when does an owner consistently override a playbook step? That step is wrong. Evaluator-rejection clusters: which deliverables get rejected the same way three times running? That brief is wrong.
How the loop closes
Lessons land in a typed memory; the engineering team reviews weekly and folds the high-signal ones into the next release as prompt edits, playbook revisions, or new circuit-breaker rules. Owners see the change in release notes and a meta-loop tag in the lesson it resolves.