# What is Tool Calling?

Also called tool use, agent tool use.

Tool calling is the pattern in which a language model is given a set of external capabilities, chooses one, has the surrounding application run it, reads the result, and repeats until the task is finished. It is largely synonymous with function calling, with different vendors preferring different names for the same underlying mechanism.

A tool calling loop has four steps that repeat. The model receives the conversation plus the tool descriptions, it either answers or requests a tool, the application executes that tool and returns its output, and the model reads the output and decides what comes next. The loop ends when the model produces a final answer or a limit on steps, time, or cost is reached.

Vendors settled on different words for this. Some documentation says function calling, some says tool use, and some says actions, but the mechanics are the same. The word tool became preferred as capabilities grew beyond ordinary functions to include retrieval, code execution, browsing, and other agents, none of which read naturally as a function call.

Two failure patterns dominate in practice. Giving a model too many tools degrades selection accuracy, because descriptions compete for attention and several may look applicable to the same request. Giving it tools with vague descriptions produces confidently wrong choices. Reliability usually improves more from pruning and rewriting tool descriptions than from changing models or rewriting the system prompt.

Because each iteration costs tokens and time and can cause real side effects, production systems bound the loop. Common controls are a maximum number of steps, an allow list of tools per context, a requirement for human approval before writes and deletions, and idempotent tool design so that a retried call does not duplicate work.

## Key points

- A repeating loop of choose, execute, observe, and continue.
- Interchangeable with function calling in most documentation.
- Too many or vaguely described tools hurt selection accuracy.
- Bound the loop with step, time, and cost limits.
- Gate writing and deleting tools behind human approval.

## In practice

Asked to summarize this quarter's churn, an assistant first calls a search tool to find the churn report, then calls a file reader on the top result, then calls a calculator on the monthly figures, then writes the summary. Four tool calls, each result feeding the next decision. Had the file reader returned an error, the model would have seen that error and could have tried the second result instead.

## Related terms

- [Function Calling](/en/glossary/function-calling)
- [Tool Schema](/en/glossary/tool-schema)
- [MCP Client](/en/glossary/mcp-client)
- [Sandbox](/en/glossary/sandbox)
- [Idempotency](/en/glossary/idempotency)

[Back to the AI Glossary](/en/glossary)