AI Engineering Philosophy

Status: Canonical
Last updated: 2026-07-05

---

1. Execution is centralized — PL-003 owns agent runtime, tools, planners
2. Learning is centralized — PL-004 owns version registry, recommendations, optimization
3. Products compose — adapters, UI, domain tools, capability registration
4. Evaluation is isolated — PL-006 benchmark/acceptance never changes production paths
5. Human approval is mandatory — no silent model/prompt promotion
6. Closed loop — measure → accept → register → deploy

See AI/README.md and ADR/ADR-007-benchmark-never-changes-runtime.md.

Breadcrumbs: Home → Engineering → AI Philosophy