Tool Use (LLM)

Tool use in language models refers to any pattern in which the model delegates part of a task to an external system rather than producing the answer purely from parametric memory. Typical tools include calculators, web search, code interpreters, vector stores, calendar and email APIs, and domain-specific services. The model's role is to decide when a tool is appropriate, choose which tool to use, and format the call; the surrounding runtime executes the tool and returns the result. An early influential demonstration was Toolformer (Schick et al., 2023), which showed that a base model could be fine-tuned in a self-supervised way to insert API calls for a calculator, question-answering system, search engines, a translator, and a calendar directly into its generations, substantially improving zero-shot accuracy on downstream tasks. Production deployments now usually rely on Function Calling (LLM) interfaces and ReAct prompting-style loops to orchestrate multi-step tool use. Tool use is the immediate context in which capability hallucination arises: the more naturally a model talks about taking actions, the easier it is for it to do so even when no tool is actually wired up. Robust agent design therefore treats the registered tool list as the single source of truth about what the agent can do, and validates every claimed action against an observable tool response.