The ability of an LLM to invoke external functions (APIs, tools) in a structured way, allowing the model to execute real-world actions.
Detailed explanation
Tool calling (or function calling) is the mechanism by which an LLM can decide to call an external function — such as an API, database, or script — as part of its response. Instead of just generating text, the model produces a data structure that specifies which function to call and with which parameters. The problem observed in production: better models sometimes get worse at tool calling. Opus 4.8 and Sonnet 5 invent fields when calling tools outside the Claude Code standard, creating integration risk for those building on the Anthropic API. For product managers, tool calling is what transforms an LLM from a text generator into an agent capable of executing actions, but it requires rigorous output validation.
How to use it in product decisions
Treat every tool call as a software operation exposed to invalid input, repetition, and side effects. Use narrow schemas, server-side validation, idempotency, and explicit confirmation for irreversible actions. Do not trust tool selection or arguments merely because the JSON is valid. Record intent, validated parameters, result, and error in an auditable form. Measure correct calls, retries, safe refusals, and failure impact. The model proposes the action; the application remains responsible for authorization and execution.