The signal I watch for is the first one on your list, small changes creating unrelated regressions. That is the point where prompt-and-test stops being cheap, because every change needs a human to rebuild confidence that the tooling should be supplying. We've kept the agent and made the gates deterministic instead, tests and static checks that run on every change regardless of who or what wrote it. The cost of change stays visible on the build, which is most of what the hybrid model needs.