A bad deploy is legible. The alert fires, the rollback runs, blast radius is the customer base, recovery is measured in minutes. That's why two-week cycles worked. Failure was cheap because failure was loud.

A bad agentic pattern is illegible.

The PR passes lint and tests. The reviewer agent approves. The diff merges. Six months later you're untangling architectural drift you can't trace to a single decision, because no single decision caused it. It accumulated across hundreds of PRs that each looked locally fine.

The cheap-failure doctrine that made the old SDLC work assumed failure was detectable. That assumption is what breaks.

Most teams building the agent-PR loop got the throughput right. They didn't replace the part that always asked: would we notice if this was wrong?

The loop runs faster. It doesn't see better.

Full piece, with what the new detection layer actually has to do: