Large language model agents increasingly need to work with tool libraries that may hold thousands of options. As new needs arise during a task, the agent must find which tools are relevant. The challenge is not just retrieval but deciding among multiple plausible candidates.

A new arXiv preprint, "Explore Separately, Adopt Selectively: Execution-Guided Tool Use over Large Tool Libraries," describes an approach that uses execution to inform this decision. The abstract notes that when several candidates could satisfy the same need, execution may be required to determine which one actually works. The method appears to separate the exploration of tools from their adoption, using execution results to selectively adopt the most suitable ones.

The source is a single preprint abstract, so details on the method's implementation or evaluation are not yet available. The work points to a practical direction: making tool selection in LLM agents more robust by grounding it in actual tool behavior rather than relying solely on static descriptions or rankings.