Developer finds AI-assisted test suite missed real bug due to flawed test fixture

A solo developer building a co-op car theft game in Unity 6 set up a three-agent AI loop — Architect, Builder, and Critic — to design, implement, and rigorously review code through mutation testing. The Critic agent was required to prove test failures by actually mutating production code and watching assertions break, not merely flagging weak-looking tests. This process appeared to confirm a multiplayer bug fix was working, with 161 out of 161 tests passing and the targeted mutation correctly causing a failure. However, the bug persisted in production because the test fixture used a stub that always reported server authority as true, while the real networking library FishNet de-initialises objects before destroying them, meaning the authority check silently returned false on every peer. The incident highlights a key limitation of mutation testing: it can verify that a test is load-bearing for what it asserts, but cannot detect when the test fixture itself fails to model real production conditions.