What is tool use?

Tool use is an AI model calling capabilities outside itself, such as a search, a calculator, a database query or an action in a business system, to get information or get something done. It is what turns a model that talks into an agent that works. Every tool an agent can call is a route into your systems, and every call is a decision.

Why does tool use matter once agents act?

Because tools are where the consequences are. A model with no tools can only produce text. A model with a "send email" tool can contact your customers. With an "update record" tool it can change your data; with a "create payment" tool it can move your money. The set of tools defines the most an agent could ever do, before any other control applies.

That is why we treat every tool call as a decision. Reading a policy and refunding a customer are both tool calls, but they deserve very different scrutiny.

How does tool use work in practice?

Each tool is described to the model: its name, purpose and inputs. During a task the model chooses a tool and supplies arguments, usually through function calling. The application runs the tool and returns the result, which the model reads before its next step. A walk-through of the full cycle is in how an AI agent uses tools.

Three practices keep tool use safe:

• Narrow tools. "Issue a credit up to the account's limit" is safer than "write to the billing database".
• Agent identity, not borrowed logins. The agent connects as itself, with its own credentials and permissions. Whether it should ever act through someone else's account is covered in can an AI agent use a person's login?
• A check before execution. Each call goes to a control layer that applies the agent's bounds and returns a verdict.

New tools should go through a request and review, not be added in code by whoever needs them; a tool connection request template gives the fields to capture.

What does that look like in an example?

A finance team's reporting agent has two tools: a read-only query on the BI platform and a "publish report" tool that posts to a shared channel. Asked for the month's revenue by region, it queries the data, builds the summary and proposes publishing it. The query runs immediately. Publishing is held for the finance manager, because the agent's certificate covers drafting reports, not releasing them. Setting up the first tool safely is covered in how to connect an AI agent to your BI tool.

How is tool use different from function calling?

Function calling is the mechanism: the model returns a structured request naming a function and its arguments. Tool use is the broader idea of an agent using external capabilities, which may be implemented through function calling, through generated code, through a browser, or through another agent. Function calling is one way to do tool use, and the most common.

Where Pryme fits

In Pryme, connections are listed in the agent's brief, and each tool call is checked against the minimum chain before it runs. The verdict, and the bound that decided it, is filed. A tool not listed in the brief cannot be called at all. See the governed agentic operating system, explained.