Tool use

The capability of a large language model to invoke external functions, APIs, or services as part of its reasoning loop. Tool use turns a stateless text-in-text-out model into an agent: the model decides which tool to call, structures the call's parameters, reads the tool's response, and incorporates the response into its next decision. The same capability is shipped under the names 'function calling' (OpenAI), 'tool use' (Anthropic), and 'function declarations' (Google) — the protocols differ but the operating concept is the same.

Tool use is the architectural primitive that distinguishes an agent from an assistant. Every governance-relevant question — what permissions does the agent have, what audit substrate captures decisions, what error class fires when something breaks — collapses to questions about the tool surface. Procurement teams that buy an agent without enumerating its tool surface inherit a permission boundary they didn't review. The reference protocol layer for vendor-portable tool use is the Model Context Protocol (MCP); see the MCP entry for the standardisation track.

Related frameworks

• MTTD-for-Agents →

Articles that analyse this term

• MCP and the coming standard for enterprise agent tooling
• AI assistant vs AI agent: the procurement distinction
• Multi-agent architecture playbook for enterprise AI

Primary sources

• Anthropic. Tool use with Claude (API documentation)
• OpenAI. Function calling (API documentation)