Does tool selection accuracy get worse as an agent gets more tools?

Yes. The more tools a model has to choose from in a single call, the worse its accuracy at picking the right one gets, because it is weighing more candidate descriptions against the same request. A 2026 study on adaptive tool lists found that on medium-difficulty queries, Claude Sonnet scored 60.9% tool-selection accuracy when handed a fixed list of five tools, against 76.8% when the list was narrowed first to only the tools relevant to that query, a 16-point gap from tool count alone. The effect holds even when the right tool is present in the list; more candidates just means more ways to pick wrong.

The practical response is not fewer tools overall, it is fewer tools presented per call. Retrieving or filtering the tool list down to what a given request plausibly needs, then handing the model that shorter list, recovers most of the accuracy a large tool catalog costs.