Tool Use & Function Calling

The 2023-2024 productisation of LLM tool use turned ReAct and Toolformer's prompting recipes into API-level features. OpenAI's Function Calling (June 2023), Anthropic's Tool Use (April 2024), and Google's Gemini function-calling are all variations on the same pattern: the developer registers a list of tool schemas; the model emits a structured tool call when appropriate; the application executes the call and feeds the result back. Tool use is now infrastructure for any non-trivial LLM application.

The protocol

A typical function-calling exchange has four roles:

• Developer sends the user prompt plus a list of tool schemas (JSON Schema for arguments, plus name and description).
• Model either answers directly or emits a structured tool-call message:

{
  "tool": "search",
  "arguments": {"query": "OpenAI revenue 2024"}
}

• Application executes the call, returns the result as a "tool message" appended to the conversation.
• Model continues, conditioning on the tool result, possibly calling more tools or producing the final answer.

The provider's training has made the model output well-formed JSON tool calls reliably. The application doesn't need to parse from natural language; the API gives a structured object.

Why function calling won

Pre-2023 tool use was prompt engineering: write a careful system prompt explaining the tool format, hope the model follows it. Failure modes were endemic — malformed JSON, hallucinated tool names, wrong argument types. Function-calling APIs solved these by:

• Constrained decoding — the inference engine ensures outputs match the JSON schema (sometimes via grammar-constrained sampling, sometimes by training).
• Standard formats — tool_calls and tool message types in the OpenAI API became the de-facto cross-provider standard.
• Validation at the boundary — providers reject malformed tool calls before they reach application code.

The combination made tool use predictable enough to build production systems on, which is what kicked off the 2024 agent-product wave.

Tool-use training

Frontier models are trained for tool use, not just prompted. The training mix typically includes:

• Synthetic tool-use trajectories — generated from a strong teacher model executing on tool sets, then filtered for correctness.
• Human-collected tool-use examples — paid labellers writing realistic queries and ideal tool sequences.
• Negative examples — cases where the model should not call a tool (the answer is already known).
• Multi-tool composition — sequences requiring multiple calls in correct order.

Anthropic's Tool Use, OpenAI's Function Calling, and Google's Gemini all use some variant of this recipe. The training data is closely-held but the format is standardised.

What's a good tool

The 2023-2024 deployment experience surfaced principles for tool design:

• Atomic. Each tool does one thing. "search" + "fetch_url" + "summarize" beats one mega-"web_research" tool.
• Composable. Tool outputs in one call should feed naturally into the next call's arguments. JSON-typed outputs make this work.
• Idempotent where possible. A tool the model can retry without causing harm is much easier to use safely.
• Well-described. The tool's docstring is what the model sees. Vague descriptions cause wrong-tool-for-the-job errors.
• Schema-tight. Loose argument schemas invite the model to send malformed calls. Tighten them where you can.

These principles are now standard SDK guidance.

Provider features by 2024

The 2024 generation of LLM provider APIs supports:

• Parallel tool calls — one assistant turn can emit multiple tool calls to be executed concurrently.
• Multi-turn tool use — the model holds state across many turns of (call, response) without re-prompting from scratch.
• Forced tool calls — the developer can require the model to call a specific tool, useful for orchestration.
• Strict mode / structured outputs — the model is guaranteed to emit a valid JSON Schema-conforming response.
• Computer use (Anthropic, late 2024; OpenAI Operator, 2025) — tools that include "click", "type", "screenshot" for browser/computer control. See computer use.

What tool use enables

The 2023-2024 tool-use wave produced:

• Coding agents — Cursor, Cody, Devin, Aider, Codex CLI, Claude Code. Tool use of code editing, terminal commands, search, file reading.
• Research agents — Perplexity Pro, ChatGPT with browsing/search, Claude with web search. Tool use of search engines, content fetchers, citation formatters.
• Workflow automation — Zapier AI, n8n, custom automations. Tool use of CRM APIs, calendars, databases.
• Robotics / IoT — early-stage. Tool use of physical-world actuators with safety wrappers.

By late 2024, "what tools should we expose to the model?" had replaced "what can the model do?" as the central practical design question.

What to read next

• ReAct & Toolformer — the conceptual ancestors.
• Agent Frameworks — orchestration on top of tool-use APIs.
• Coding Agents — the most-deployed downstream application.