AI Infrastructure

Tool calling

Tool calling is the capability that lets a language model invoke a function, API, or MCP server to fetch data or take action.

Tool calling is the mechanism that turns a chat model into an agent. The model produces a structured request naming a tool and passing arguments. The harness runs the tool and returns the result. The model then decides what to do next.

Every modern frontier model supports tool calling. The capability itself is commoditized. What separates working agents from broken ones is the quality of the tool definitions: descriptions the model can actually reason about, schemas that constrain inputs cleanly, errors that teach the model how to recover, and return payloads the model can use without extra parsing.

Tool calling sits at the core of every useful agent pattern: retrieval, action execution, multi-step planning, and handoff to other agents. A team that cannot get tool calling right cannot ship agents that work in production.

The Amdahl view

Teams that treat tool calling as a free capability are in for a rough time. Teams that treat tool descriptions, schema design, and error handling as first-class engineering are the ones whose agents actually work in production. Amdahl's MCP was designed tool-call first, with every action shaped around what an agent can reason about, not what a developer finds convenient to expose.

See customer intelligence running on your own customer conversations.