Verifier Rejected No Answers, Yet Agent Score Dropped; Search Tool Blamed

An agent-patterns project added a verify-before-finish gate to a ReAct agent, and on 15 two-hop questions with Gemini 3.1 Flash-Lite, plain ReAct solved 10 while the gated version solved 9. The author first attributed the drop to the verification step consuming iteration budget, but a later replay showed the verifier accepted all 11 of its answers and never rejected one. Replaying cached responses offline showed the two variants diverged only through the one added system-prompt sentence, which shifted search behavior on a few tasks. The main failure source was the search tool's fuzzy matching, which fails on phrasings like "Lake Azurine" that a stored key like "azurine lake" doesn't match. Of 125 searches across the Gemini runs, 64 returned no results, and the agent made no lookup calls.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.


Discussion (0)
Log in to join the discussion and vote.
Log in