Ask HN: How do you define "done" for long-running AI agents?

[IntelliAvatar]

I've been working on long-running automation / agent systems, and one thing I keep running into is how hard it is to define "done".

Demos are easy: a task finishes once the happy path works. Real systems are messier — partial failures, retries, idempotency, unclear terminal states, and the question of when to stop or escalate.

For people who’ve built schedulers, agents, or other long-running systems: how do you define "done" in practice? Is it a state machine, invariants, timeouts, external signals, or just operational heuristics?