← sources

A tool returning success is not evidence — read the result back from the system that owns it

source: reddit (reddit:AI_Agents)published 2026-11-03 05:54ingested 2026-11-04 02:36

retained text

Most agent failures I see discussed are about bad reasoning. The quieter one is about bad evidence: the agent's tool returned success, so the agent reported success, and nobody checked the system that actually owns the result. A tool success usually means "accepted", "queued", or "no error raised" — not "the change landed in the state you intended". Partial writes, async queues, and defaults that silently substitute for missing fields all pass a tool-level check while leaving the destination wrong. The pattern I'd use instead: treat every consequential write as unverified until a separate read from the system of record confirms it. Not the tool's response, not the agent's log — a fresh read of the record, in the state you expected, under the ID you expected. If the read can't run yet, the task is deferred, not done, with a recheck time and an owner. Two failure modes worth naming. First, re-reading through the same wrapper that wrote the data proves nothing; read through a different path if you can. Second, a read-back that only checks "a record exists" will pass on a partial write — check the fields that matter, or at least a sample of them. Where do you draw the line on what deserves a read-back? Every write, or only the irreversible and externally visible ones?

document fetched 2026-11-04 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xml1y8/a_tool_returning_success_is_not_evidence_read_the