← sources

When your agent can't read the result back, what counts as proof?

source: reddit (reddit:AI_Agents)published 2026-10-26 06:10ingested 2026-10-27 02:35

retained text

Most agent verification advice assumes you can read the destination state back. A lot of real actions break that assumption: an email sent, a webhook fired, a ticket opened in someone else's system, a payment submitted. The agent reports success, and there is no record you can open to check it against.

I'd grade the evidence instead of treating the agent's report as proof.

First, a provider receipt means 'accepted', not 'done'. An API returning 200 for a send call proves the request was received. It does not prove the message arrived, was processed, or had the intended effect. Wording the claim to match the evidence matters: 'accepted and logged' is honest, 'sent and delivered' is not, unless you have the delivery record.

Second, where you can, prove delivery out of band. In staging, send to an inbox you control and assert on the delivered message, not the send call. Point staging webhooks at a capture endpoint and compare the payload you received with the one you meant to send. In production, a queryable mail log, a BCC copy, or a delivery record from the destination system gives you an observable that is independent of the agent's transcript.

Third, when the destination is genuinely unreadable, don't close the item on a receipt alone. Preserve three things together: the target frozen before the action ran, the receipt, and a timestamp. Then make the item a typed obligation — what evidence is still missing, who owns it, what retriggers the check, and when it expires — instead of marking it done and moving on.

The uncomfortable part is that this still leaves a gap for low-observability actions. You can narrow it, but you can't always close it, and pretending otherwise is how 'done' stops meaning anything.

For actions you run in production, what's the strongest out-of-band check you actually use — and where do you draw the line between 'accepted is good enough' and 'I need real proof'?

document fetched 2026-10-27 00:20 · canonical https://reddit.com/r/AI_Agents/comments/1xfoyk0/when_your_agent_cant_read_the_result_back_what