A tool call is the action step that turns an assistant from talking about work into doing part of it. We often describe tools as giving the AI arms and legs: the model is the brain, and the tool is what lets it reach into a file system, query a database, or call an API. The tool call is the actual reach.
Think about an office assistant at a desk. If you ask which invoices are still unpaid, they can’t answer from memory. They open QuickBooks, run the report, read the result, and tell you what they found. Opening QuickBooks and running that report is the tool call.
How it shows up
In agentic work, tool calls happen constantly. Codex reads files, Claude Code runs tests, an agent queries a CRM, a browser tool clicks a button. Each moment has three parts: the agent chooses the tool, sends the input, and receives an output it can use next. This is closely related to function calling, which is the structured way a model asks for a specific action. A tool call is the broader working moment you see in practice.
Why you care
Tool calls are where consequences enter the room. A text answer can be wrong, but a tool call can change a file, spend money, send a message, or expose private information. That’s why permissions, review, and clear instructions matter. A good agent uses the right tool, reads the result carefully, and adjusts. A bad one sprays tool calls around because it’s confused.