Tool Selection

Tool selection studies how agents decide whether external help is needed, which API, model, or physical tool fits, and what subset of options to expose at each step. Common approaches include semantic and behavior-aligned retrieval, learned rankers and reinforcement-learning policies, causal precondition-effect filtering, and signals from model activations or attention; some systems also optimize tool descriptions and multi-tool plans. Evaluations use function-calling benchmarks such as BFCL, ToolBench, and τ-bench, alongside multimodal, coding, and enterprise settings, measuring task success, unnecessary calls, token cost, and robustness to distractors. A central challenge is that relevance is not the same as necessity or reliable execution: growing tool catalogs, ambiguous goals, and interactions among tools make it hard to generalize from fixed-menu tests to real agent workflows.