What Is Tool Calling in LLMs? How Agents Act
A model that can only write text is a chatbot. A model that can call tools is an agent. Tool calling is the mechanism that makes the difference.
- Published
How it works
- You describe tools to the model: a name, a description and the parameters each takes.
- The model, reading the conversation, decides a tool would help and outputs a structured request: which tool, with which arguments.
- Your code — not the model — runs the tool, e.g. looks up an order.
- The result goes back to the model, which continues: answers, or calls another tool.
The model never touches your systems directly. Your code decides whether to execute each request, which is where permissions and approvals live.
Designing tools that work
- Few, clear tools beat many overlapping ones. The model picks better from ten well-named tools than forty similar ones.
- Descriptions are prompts: say when to use the tool and when not to.
- Return concise, structured results — not a raw 5,000-line API response.
- Make errors informative: "order not found, check the number format" lets the model recover.
- Make risky tools require confirmation or human approval.
Where MCP fits
The Model Context Protocol standardises how tools are described and served, so one tool server can be used by many AI applications. If you are building tools that several agents or clients will use, packaging them as an MCP server saves repeated integration work.
Frequently asked questions
Can the model call tools I did not give it?
No. It can only request tools you defined, and your code decides whether to run them.
How many tools is too many?
Accuracy tends to drop as overlapping tools pile up. Group rarely used tools or load them only when relevant.
Is tool calling reliable?
With clear tool design and evals, very. Most failures trace back to vague descriptions or messy outputs.
