On its own, a language model can’t act on anything outside itself; it only produces output, usually text. Tool use closes the gap to the real world: the model is told what functions it can call and, when it decides one is needed, it outputs a structured request — search_orders(customer_id) — that your code executes and hands the result back for the model to use. The model chooses which tool and what arguments; your system runs it.
This is the mechanism that turns a model into an agent. Planning is useless without the ability to act, and tool use is that ability — reading a file, querying a database, calling an API, triggering a workflow. The reliability questions that dominate agent-building (does it pick the right tool? pass valid arguments? recover when a call fails?) are all tool-use questions. Open standards like the Model Context Protocol exist to make the same tools work across models and apps without bespoke wiring each time.
A valid call is not an allowed one
An illustrative check: an agent asks an order-lookup tool for an order ID that belongs to a different customer. The request matches the tool’s schema perfectly, so validation passes; the service that holds the orders still has to check that the authenticated user the agent is acting for may see that order before returning it. A second case: a write times out. If the API supports idempotency keys and the original request included one, retry with that same key and the same parameters. Otherwise, confirm the operation’s status before deciding whether to retry, or you risk doing it twice. Schema validation establishes neither permission to act nor whether a write took effect.
Sources
- Anthropic: Tool use with Claude — Explains that the model decides when to call a tool from its description and returns a structured call that your application runs, or that the provider runs for server tools; checked October 25, 2026.
- OpenAI: Function calling — Describes function tools defined by a JSON schema that let a model connect to data and actions your application provides; checked October 25, 2026.
Go deeper
- How does an AI agent decide which tool to use? AI, decoded · How AI Agents Use Tools
- How do AI voice agents work? AI, decoded · How AI Voice Agents Work
- What makes an AI agent different from an LLM? AI, decoded · 4 Things That Turn a Model Into an Agent
- Anthropic: Writing effective tools for agents article
Improve tool names, descriptions and response formats, then test whether agents choose and call them correctly.
From the conversation
-
Beyond Transformers: How Liquid AI Is Rethinking LLM Architecture | Maxime Labonne -
The Making of Gemini 2.0: DeepMind's Approach to AI Development and Deployment | Logan Kilpatrick -
GenAI Predictions for 2025 | Databricks & Cohere -
This is the worst AI tutor we'll ever have | DataCamp -
Inside IBM's watsonx: Building Enterprise AI That Ships | Dr. Maryam Ashoori -
Agents Are Overtaking Humans in Search: Why Search Must Be Rebuilt -
Most of the Web Will Never Get APIs for AI Agents | Dhruv Batra, Yutori -
Every AI Agent Has an Evaluation Gap | Alex Ratner, Snorkel AI