Tool calling (also called function calling) lets a large language model use external functions. The application describes each tool with a JSON schema; the model answers with a structured request to call one; the application runs it and returns the result. This is the core mechanism behind Agentic AI.
Loop
Steps:
- the application sends the conversation plus a list of tools, each with a name, a description and a JSON schema for its arguments
- the model replies with text, or with one or more tool calls: a tool name and arguments that follow the schema
- the application validates each call and runs it
- the application sends back a tool result for every call, linked by its call id
- the model reads the results and decides the next call, asks a question, or gives a final answer
- repeat until done, with a limit on the number of turns
Example Schema
{
"name": "pick",
"description": "Grasp the object at a position and lift it.",
"input_schema": {
"type": "object",
"properties": {
"xyz": {"type": "array", "items": {"type": "number"},
"minItems": 3, "maxItems": 3}
},
"required": ["xyz"]
}
}Model Does Not Run Anything
- it only produces a request, and the application decides whether to run it
- so every call must be checked: the tool exists, the arguments are allowed and complete, the types are right, and values are safe (for example a position inside a robot's workspace)
- a rejected call can be returned with the reason, so the model can correct it
Design Tips
- write descriptions that say when to use and when not to use each tool
- a "question" tool (for example
ask_user) lets the model ask instead of guessing when a request is ambiguous - keep schemas and the real function signatures in sync, and check this at start-up
- return useful results ("position below table by 3 cm"), not just "failed", so the model can recover
- a strict schema mode guarantees that calls match the schema
- forcing a structured answer through a single tool is a reliable way to get JSON output, for example a list of object names
Risks
- the model can invent arguments, such as coordinates of objects it was never told about, so it must be told to use only given values
- long loops cost time and money, so cap the turns