Tool Use & Function Calling

Tool Use (also called function calling) is the capability that allows language models to interact with external systems — databases, APIs, code interpreters, web browsers, and physical devices — by generating structured calls that trigger real-world actions and return results the model can reason about. This is the foundational capability that transforms a language model from a text generator into an agent capable of doing things, not just saying things.

The mechanics are straightforward but profound. Instead of only producing natural language, a model trained for tool use can output structured data (typically JSON) that conforms to a predefined schema — essentially calling a function with specific arguments. The system executes that function, returns the result, and the model incorporates the result into its ongoing reasoning. A model asked "What's the weather in Tokyo?" doesn't hallucinate an answer — it calls a weather API with {"location": "Tokyo"}, receives real-time data, and reports it. A model asked to "book the cheapest flight to London next Tuesday" can search flight APIs, compare prices, and initiate a purchase. The shift from generating text about actions to generating the actions themselves is what makes agentic AI possible.

The standardization landscape in 2025-2026 has converged around several approaches. Anthropic's Model Context Protocol (MCP) provides an open standard for connecting AI models to external tools and data sources through a unified protocol. OpenAI's function calling API defines tool schemas that models can invoke. Google's Gemini supports native tool use with grounding in Google Search and custom functions. The emergence of standards like MCP matters because it means tools written once can work across multiple models — creating an ecosystem of reusable AI capabilities analogous to the app store model in mobile computing.

The capability hierarchy matters. Basic tool use is single-turn: model calls one function, gets a result, responds. Advanced tool use is multi-turn: model chains multiple function calls, using the output of one as input to the next. Autonomous tool use (the agentic frontier) involves the model deciding which tools to use, when to use them, and how to recover when a tool call fails — all without human intervention. The progression from basic to autonomous tool use tracks directly with the progression from chatbot to copilot to fully autonomous agent, and models in 2026 are increasingly capable at the autonomous end of this spectrum.

Tool use also reshapes the economics of AI. A model without tools must encode all knowledge in its parameters — which means massive training costs to cover every possible domain. A model with tools can be smaller and cheaper because it can look things up rather than memorizing them. This connects directly to RAG (which is essentially tool-use for document retrieval) and to the broader trend of small language models augmented with tool access outperforming larger models on practical tasks. The agentic commerce vision — AI agents negotiating, purchasing, and transacting on behalf of humans — is fundamentally a tool-use story.