Glossary · Agentic
Tool use
Tool use is the mechanism by which an AI model calls external functions, APIs or data sources during a task, instead of answering only from what it learned in training.
Tool use, or function calling, is how an AI model invokes external functions, APIs or data sources during a task rather than answering only from its training data. The model chooses a tool and its arguments; your code runs it and returns the result. It is what lets an agent act, not just describe.
- Tool use lets a model act (query data, call an API, run code) not just generate text.
- The model decides which tool and what arguments; your code executes it and returns the result.
- It is the foundation of agentic AI: no tool use, no acting on the world.
- Every tool is arbitrary code execution and must be treated as a security boundary.
- Tool selection and argument errors are a real failure mode and must be evaluated, not assumed.
Also known as: function calling, tool calling
Tool use is the mechanism by which an AI model calls external functions, APIs or data sources during a task, instead of answering only from what it learned in training.
The model does not run anything itself. It returns a structured request naming a tool and its arguments; your code decides whether to execute it, runs it, and hands the result back for the model to read and act on.
How tool use works
You give the model a set of tools, each with a name, a description of what it does, and a schema for its inputs. During a task the model can respond in two ways: with an answer, or with a request to call one of those tools. When it requests a call, it produces the tool name and a set of arguments that match the schema.
Control then returns to your application. Your code validates the request, checks permissions, runs the tool (a database query, an API call, a code execution, a retrieval step), and returns the output. The model reads that output and continues, often calling further tools, until the task is done. This request-execute-return cycle is the atom that agentic AI is built from.
Why tool use matters for enterprise AI adoption
Tool use is the line between a model that describes work and a model that does it. A language model on its own can draft an email about a shipment; with tools it can look up the shipment, check the exception, and update the record. Everything an enterprise actually wants an agent to accomplish runs through a tool call.
It is also where the security boundary sits. Because a tool call is, in effect, a request to run code against your systems, it is the point where permissions, validation and (for consequential actions) human approval must be enforced. The MCP specification is explicit that tool descriptions from an untrusted server should be treated with caution and that a host should obtain user consent before invoking a tool (Model Context Protocol specification). Getting tool use right is therefore both what makes agents useful and what keeps them safe.
Common mistakes with tool use
The first mistake is exposing too many tools with overlapping descriptions. The model has to choose, and ambiguous or near-duplicate tools produce wrong-tool errors that look like model failures but are really design failures. Fewer, clearly distinct tools per task work better.
The second is trusting the arguments. A model can produce a well-formed call with a wrong value, so validate inputs before executing and never let a tool call bypass the checks a human action would face. The third is not measuring it: tool selection and argument accuracy belong in an evaluation harness, because you cannot fix an error rate you never counted.
Related terms
- Agent orchestration: coordinating many tool calls into a reliable task.
- Model Context Protocol: an open standard for exposing tools to models uniformly.
- Agentic AI: the systems that use tools to act, not just answer.
- Evaluation harness: how tool selection and argument accuracy get scored.
How Chokmah approaches tool use
We treat every tool as a security boundary, not a convenience. In a workflow sprint we keep the tool set for each task small and unambiguous, validate arguments in our own code before anything runs, and put a human checkpoint in front of any tool that changes a record with real consequences. Tool selection accuracy goes into the evaluation harness from day one, so the client can see the error rate rather than assume it away.
Sources
- Model Context Protocol, Specification (2025-06-18). https://modelcontextprotocol.io/specification/2025-06-18
Related terms
- Agent orchestrationAgent orchestration is the coordination layer that routes work across multiple AI agents and tools so a multi-step task completes reliably, handling state, retries, handoffs and stopping conditions.
- Model Context Protocol (MCP)The Model Context Protocol (MCP) is an open standard that defines how AI applications connect to external tools and data through one uniform interface instead of many bespoke integrations.
- Agentic AIAgentic AI is software that uses a language model to plan and carry out multi-step tasks by calling tools, observing the results, and choosing its next action in a loop.
- Evaluation harnessAn agent evaluation harness is a repeatable test suite that scores an AI agent's outputs against fixed, versioned cases before and after every change, so teams can tell regression from variance.
Frequently asked questions
They describe the same mechanism. Function calling is the term most model providers use for the interface: the model returns a structured request naming a function and its arguments. Tool use is the broader concept: the model reaching outside itself to act, whether the tool is an API, a database query, a code runner, or a retrieval step. In practice the words are interchangeable.
No, and this matters for security. The model only proposes which tool to call and with what arguments. Your own code decides whether to run it, executes it, and returns the result for the model to read. That boundary is where you enforce permissions, validation and human approval. Treating a tool call as arbitrary code execution (because that is what it can be) is the correct security posture.
Tool use is the capability; an agent is what you build with it. An agent runs a loop that plans, calls tools, reads the results, and decides the next step. Without tool use an agent can only produce text; with it, the agent can act: retrieve a document, update a record, send a message. Tool use is therefore the foundation the whole agentic pattern stands on.
Usually because tool descriptions are ambiguous, overlapping, or too numerous, so the model cannot tell which one fits. Wrong-tool and wrong-argument errors are a measurable failure mode, not a rare edge case, which is why tool selection belongs in your evaluation harness. Clear, distinct tool descriptions, a smaller tool set per task, and scored test cases reduce the error rate far more than a better prompt does.
Put the concept to work
We install working agentic workflows, not vocabulary. Book a free AI Reality Check.