← sources

How much verification do you require before an agent's task counts as done?

source: reddit (reddit:AI_Agents)published 2026-11-13 07:16ingested 2026-11-14 02:35

retained text

i have a few agents that publish on my behalf across a handful of channels, and what keeps biting me isn't the model getting the answer wrong. it's the task getting marked done too early.

last week one of them came back with a successful upload and a real id. thirty-six seconds later the item was showing as unavailable on the platform side, and the run had already been written into the log as complete. i only caught the pattern when i compared the log against the platform listing, because five of seven attempts had failed the same way and every single one had looked fine at the moment it returned.

so now done means three separate things: the write comes back with an id, an independent read confirms the platform agrees, and a person can actually open the thing. any one missing and the run is a failure. it costs a few more calls per task and it caught two more bad runs in the same week.

where i'm stuck is the lower end. a read-only lookup clearly doesn't need three checks. but for anything that writes something other people can see, i've stopped trusting a single success response, and i'm not sure that's discipline or just scar tissue from one bad week.

if you run agents that touch systems outside your own, do you set the bar by how visible the action is, by how hard it is to undo, or do you verify everything the same way and eat the cost?

document fetched 2026-11-14 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xv1d8k/how_much_verification_do_you_require_before_an