Tool Use
When the model calls external functions (search, calculator, DB, API) during a response · the building block for agents.
When the model calls external functions (search, calculator, DB, API) during a response · the building block for agents.
Basic
Tool use lets the AI ask the outside world for help. The model generates a structured tool call (e.g., "search(\"AI news 2026\")"), a harness executes the call, and the result comes back into the conversation. Every modern frontier API supports tool use via function calling · OpenAI, Anthropic, Google, Mistral. MCP extends this with a standardized protocol for tool providers.
Deep
Flow: model generates a JSON tool call matching a schema → harness validates + executes the call → result is appended to the conversation → model continues. Parallel tool calls (invoke multiple at once) reduce latency. Tool descriptions matter enormously · a well-named and well-described tool is called reliably; a poorly-described one is ignored. Common tools: web search, calculator, database query, code execution, file operations, API calls to external services, MCP-exposed capabilities. Safety: destructive tools (file delete, send email) should require user approval.
Expert
Function calling APIs standardized the JSON schema for tools. Parallel tool use reduces round-trips · a frontier model can emit 3-10 tool calls in a single turn. Tool-result interleaving: results are fed back as special message turns. Few-shot examples of good tool use in the system prompt improve calling accuracy 10-30%. Agent frameworks (LangGraph, CrewAI, OpenAI Agent SDK) wrap tool use with orchestration, memory, and error handling. Evaluation: toolbench (Tencent), APEX-Agents (Epoch), BFCL (Berkeley function calling leaderboard).
Depending on why you're here
- ·JSON schema for tool definitions
- ·Parallel tool calls reduce agent round-trips 2-5×
- ·BFCL and APEX-Agents are the standard benchmarks
- ·Write clear tool descriptions · biggest lever for reliability
- ·Support parallel calls · agents get faster dramatically
- ·Gate destructive tools behind user approval · safety critical
- ·Tool use drives the agent revenue wave · chat plateaus, agents grow
- ·MCP winning the open-protocol layer
- ·Framework market consolidating (LangGraph, OpenAI Agent SDK)
- ·AI that can use tools · search the web, run code, send email
- ·The upgrade from "chat" to "assistant"
- ·Every agent you use internally uses tool calling
Often confused with
Function calling is the API mechanism. Tool use is the conceptual capability. MCP is the open protocol standardizing tool definitions across vendors.
Tool use is a single call. An agent uses tools in a multi-step loop with planning and memory.
Tool use turned LLMs from autocomplete into actors. Everything in 2026 agent land is built on this one capability.