Tool Use
Tool use enables agents to interact with external systems—searching the web, executing code, calling APIs—extending their capabilities beyond text generation.
Definition
Tool use refers to an agent's ability to invoke external functions or services to accomplish tasks that cannot be done through text generation alone. Tools bridge the gap between language model capabilities (reasoning, synthesis) and real-world actions (data retrieval, computation, system modification).
Common tool categories include: - Information retrieval: Search engines, databases, APIs - Computation: Code execution, calculators, data analysis - Communication: Email, messaging, notifications - System interaction: File operations, application control
Why It Matters
Tool use is what transforms a chatbot into an agent. Without tools, a language model can only reason about information in its context window. With tools, an agent can:
Access current information: Language models have training cutoffs; tools provide real-time data.
Perform precise computation: Math in language is unreliable; code execution is deterministic.
Take real action: Agents can modify external state—creating documents, sending messages, updating systems.
Scale capabilities: New tools extend agent capabilities without retraining the underlying model.
How Tool Use Works
Modern tool use typically follows this pattern:
- Tool definition: Each tool has a schema describing its name, purpose, and parameters
- Context inclusion: Tool schemas are included in the agent's context
- Selection: The agent reasons about when to use which tool
- Invocation: The agent outputs a structured tool call (typically JSON)
- Execution: The system executes the tool with provided parameters
- Result integration: Tool output is added to context for the next reasoning step
The key insight is that the agent doesn't execute tools directly—it specifies what tool to call and with what arguments. A separate system handles execution.
Common Misconceptions
"More tools are always better" Too many tools create decision complexity. Agents may choose wrong tools or combine them ineffectively. Start with essential tools, add as needed.
"Tools are reliable" Tools can fail: APIs timeout, services error, rate limits trigger. Robust agents handle tool failures gracefully.
"The model understands what tools do" Models use tools based on descriptions, not true understanding. Poor tool descriptions lead to misuse.