The boundary of 'what AI can do' is moving. The debugging agent is a useful case study. The first generation of coding agents tried to fit into existing manual debugging workflows, using the same observability data humans relied on. The result was PR slop: fixes generated from sampled, aggregated, and missing data that looked plausible and failed in production. The next generation is being built differently: session-based, full-stack, pre-correlated runtime data, with issues deduplicated before they ever reach the coding agent. That's the workflow being redesigned around what machines actually need, and not retrofitted into what humans used to do.