What it is
Tool use, also called function calling, lets a developer describe functions to Claude with a name, a description and a JSON Schema for the inputs. When Claude decides it needs one, it replies with a tool request instead of a final answer. The application runs the function and sends the result back, and Claude carries on with that result in hand.
Running that exchange in a loop until Claude has an answer is the basic shape of an agent. The model never runs anything itself. The application decides which tools exist, checks the inputs, and can refuse or ask a person first, so the developer stays in control of every side effect.
How I use it
- The agent on mervinc.com has five small tools: site search, live traffic, this month's AWS bill, build numbers and which CloudFront edge served the visitor.
- Site search is BM25 over every page of the site, cut into passages when the function is deployed, so it never answers from a stale copy.
- Every tool call, what came back and how long it took streams to the page as its own event.
- Radar, the help assistant inside Rental Pilot, searches the help guides and files a problem report only after the person agrees.
Tool use is the right approach when a model needs facts it cannot know, like live data or a company's own documents, or when it should take an action. I keep the tools small and specific so each one is easy to check, and anything with a consequence waits for a person to agree.
Questions
What is the difference between a chatbot and an agent?
An agent can call tools. Claude asks for a function, my code runs it and returns the result, and the loop repeats until there is an answer.
Can an AI assistant take actions in my business software?
Yes, through tools, but I make anything with a consequence wait for a person. Radar in Rental Pilot only files a problem report after the user says yes.