Tools, Not Talk: The Knowledge Ceiling of Biology Agents

Many agent systems for biology add conversation to a language model: a critic reviews a draft answer, or role-played experts answer and an adjudicator decides. These systems are usually compared with a single model call, so it is unclear whether the conversation helps or the extra calls do. We give each conversational scaffold three calls and compare it with three independent answers combined by majority vote, on 150 LAB-Bench and LABBench2 biology questions scored deterministically under an analysis plan fixed before evaluation. Self-critique scores below the vote for all three models, losing up to 23 points on sequence manipulation; an expert panel beats it only for a model run without reasoning. Systems succeed and fail on the same questions: the question accounts for 56% of the variance in correctness, and the system answering it for under 1%. On the questions every system fails, mostly literature and database lookups, one call with web search and a Python shell answers 47%, against at most 16% for the same model’s three-call scaffolds; web search alone reproduces the gain, and a placebo prompt without tools does not. On biology questions like these, the baseline to beat is a vote with equal calls, and accuracy improves through tool access, not deliberation.