Tool Use
Tool use is the general capability of an AI model to invoke external functions, APIs, or software utilities during its reasoning process — such as a calculator, code interpreter, web search, or database query — in order to extend what it…
Definition
Tool use is the general capability of an AI model to invoke external functions, APIs, or software utilities during its reasoning process — such as a calculator, code interpreter, web search, or database query — in order to extend what it can accurately do beyond generating text from its own trained knowledge.
Overview
Language models are inherently limited to producing outputs derived from patterns learned during training: they cannot natively execute precise arithmetic reliably at scale, access information newer than their training cutoff, browse the live web, or interact with external systems. Tool use addresses these limitations by giving the model access to a defined set of external capabilities it can invoke mid-reasoning, then incorporate the results into its continued response. The mechanism underlying tool use is typically function/tool calling: the model is shown descriptions of available tools, and when it determines a tool is needed, it emits a structured request specifying which tool to call and with what arguments; the surrounding application executes the tool and returns the result to the model. Tool use is what elevates a model from a passive text generator into an 'agent' capable of taking actions in the world — searching the web for current information, running code to verify a calculation or generate a plot, querying a database for exact records, calling a company's internal API to check an order status, or manipulating files in a codebase. Prompting strategies like ReAct (Reasoning + Acting) formalize a loop where the model alternates between reasoning about what it needs and taking a tool action to get it, observing the result, and reasoning again, iterating until it has enough information to give a final answer. Tool use is foundational to how modern coding assistants (such as Claude Code, GitHub Copilot's agent mode, and Cursor's agent mode), research assistants, and customer-support automations operate, and it is the mechanism by which retrieval-augmented generation is often implemented in agentic systems (treating retrieval itself as a callable tool rather than a fixed pipeline step). Anthropic's Model Context Protocol (MCP) has emerged as an attempt to standardize tool definitions and access across different models and applications, reducing the need for bespoke, provider-specific tool integrations. As with function calling more broadly, reliability, argument correctness, and appropriate tool selection remain active engineering challenges, particularly as the number of available tools in a system grows.
Key Concepts
- Lets a model invoke external functions, APIs, or utilities mid-reasoning
- Typically implemented via structured function/tool calling requests generated by the model
- Extends model capability beyond training-time knowledge: live data, precise computation, actions
- Formalized reasoning-and-acting loop patterns like ReAct interleave reasoning with tool actions
- Foundational to agentic systems including coding assistants and research agents
- Underlies many practical implementations of retrieval-augmented generation
- Standardized across providers via protocols like Anthropic's Model Context Protocol (MCP)
- Reliability of tool selection and argument correctness remains an active engineering challenge