Tool use
The ability of an LLM to invoke external functions — APIs, shell commands, internal services — instead of just generating text.
Tool use turns an LLM from a text generator into something that acts. The model emits a structured call (function name + JSON arguments); a wrapper executes it and feeds the result back. Loop until done.
Almost every modern agent is built on top of tool use. The interesting variation is not whether tools exist but how many, how custom, and how the agent decides which to call.
In 2026, function calling is a stable feature across all major models. The bar has moved up to "can the agent reliably choose the right tool from a 30-tool set?" — and most still struggle past 50.
Frequently asked
What is the difference between tool use and function calling?+
They are the same thing. "Function calling" is the API name OpenAI and Anthropic use for the feature. "Tool use" is the broader concept. In practice the terms are interchangeable.
How many tools can an agent reliably use?+
Frontier models in 2026 can pick correctly from 20–30 tools most of the time. Past 50 tools, accuracy drops sharply and most teams split into multiple specialized agents.