AI library

Agents

Tool use (function calling)

Last reviewed 11 October 2026

Overview

Tool use lets a model call functions you define: search a database, call an API, run code, edit files. The model does not run anything itself. It returns a structured request (tool name plus JSON arguments); your code runs it and sends back the result. This is how LLMs act in the world and the core of every agent.

Key concepts

  • Tool definition: name, description, and JSON Schema for the inputs. The description is a prompt: it is how the model knows when and how to use the tool.
  • Loop: request, then tool_use response, run the tool, send tool_result, repeat until the model gives a final answer.
  • Parallel tool calls: several calls in one turn; return all results together.
  • Strict mode: guarantees arguments match the schema exactly.
  • Server tools: tools the provider runs for you (web search, web fetch, code execution).
  • Tool search / deferred loading: with many tools, load definitions only when needed to save context.
  • Programmatic tool calling: the model writes code that calls your tools, so large intermediate results never enter its context.
  • Tool runners: SDK helpers that run the loop for you.
  • Standard connections: MCP packages tools as reusable servers.
  • Forcing a tool: tool_choice can require a tool call on many APIs, but Claude Fable 5.1, Opus 5.5, and Sonnet 5.5 reject forced tool_choice (any or tool); use auto with clear instructions, strict schemas, or structured outputs instead.
  • Hosted computer use: providers increasingly run the browser or desktop for the agent (OpenAI's Agents API since 2026-09-29), while the application handles site approvals and sign-in.

How to use it

  • Write tool descriptions like documentation for a new colleague: purpose, when to use, input meaning, examples, edge cases.
  • Prefer a few well-designed tools over many overlapping ones.
  • Return concise, high-signal results; truncate or paginate big outputs.
  • Return errors as tool results (with an error flag) so the model can recover.
  • Validate inputs and require confirmation for destructive actions.

Current state

  • 2026-10-07: Claude Haiku 5.5 added computer and browser use to Anthropic's smallest tier.
  • 2026-09-29: OpenAI's Agents API added hosted computer use.

Open questions

  • Best practices for very large tool catalogues (hundreds of tools).

Sources