Your agent said 'done' — here's how to make that word mean something

The most expensive sentence in agent systems is "task completed." Not because agents lie — because "done" usually means "I stopped without an error," which is a completely different claim from "the thing you wanted actually happened." Threads here keep circling the same pain: the agent reports success, then a human has to go check whether anything actually happened. The fix isn't more logging. It's deciding, per action, what evidence would prove the outcome — before the agent runs. Three levels of proof, cheapest first: 1. The receipt. The target system acknowledged the action with an ID you can look up later. Not the agent's own log — the other system's record. If there's no ID, there's no receipt, and "done" is just a diary entry. 2. The readback. After the action, read the target system and confirm the state changed the way you intended. A ticket exists and has the right fields. A row exists and matches what was sent. This catches the 200-that-wasn't class of failure, where the call succeeded and the effect didn't. 3. The independent check. For high-stakes actions, verify with something that doesn't share the agent's failure modes. An LLM judge reading the agent's summary is not independent — it inherits the same blind spots. A deterministic check (schema, count, balance, status transition) is. Rule of thumb: the harder an action is to reverse, the higher the level of proof it should need before anyone calls it done. A draft can live on level 1. A payment can't. What do you require before you trust a "done" — or do you just check by hand every time?