Tool-Augmented Language Model Agents

Tool-augmented language model agents decide when and how to use tools such as web search, code interpreters, and application APIs, then incorporate the results into multi-step tasks. Common approaches include retrieval and filtering of tool menus, planning and memory, structured tool-call interfaces, and supervised fine-tuning or reinforcement learning in executable environments. Benchmarks such as BFCL, τ²-bench, and AppWorld test tool selection, argument accuracy, and task completion, while newer evaluations examine long-horizon workflows and failures in stateful settings. A central challenge is making tool use reliable and efficient: tools can improve answers, but unnecessary calls, faulty outputs, and mismatches between models and agent harnesses can undermine performance.