# How do I catch agent hallucinations when my error rate stays flat?

Catch it by grading content instead of status: check whether the needed information was retrieved, whether the answer came from that context or the model's own prior, and whether it represents that context accurately. Those are three separate checks, not one score, because each points at a different fix.

Error rate misses this because it tracks whether a run crashed, timed out, or returned a bad status, not whether its answer was true. A hallucination produces a clean 200 and fluent, well-formed text, so it never trips that signal; you're watching the wrong layer of the system. Run the content checks on a real share of production traffic, not a thin manual sample, since a rare hallucination has to show up enough times to be worth trusting. An error-rate dashboard staying flat isn't evidence the agent is telling the truth; it's evidence nothing is checking.

---

Source: https://tessary.ai/answers/hallucinations/how-do-you-catch-hallucinations-when-error-rate-is-flat
More on Hallucinations: https://tessary.ai/answers/hallucinations
From Tessary, agent reliability for AI agents in production: https://tessary.ai