Tool Use

What Is Tool Use?

Tool use is the capability that allows an AI agent to reach beyond its training data and interact with the real world. Instead of relying solely on what the LLM learned during pretraining, the agent can call external functions, query APIs, run code, or browse the web to gather fresh, accurate, or specialized information.

Without tool use, an LLM is confined to what it already knows — which is static and may be outdated. With tool use, an agent becomes dynamic: it can look things up, take actions, and produce results grounded in live data.

How Tool Use Works

At a technical level, the LLM is provided a list of available tools and their input schemas. When the model decides a tool is needed, it outputs a structured request (typically JSON) specifying which tool to call and with what arguments. The agent runtime intercepts that request, executes the real function, and feeds the result back to the model as a new observation.

The steps are:

1. Tool definitions provided — The agent is initialized with a list of tools and their descriptions.
2. LLM selects a tool — Based on the task, the model outputs a tool call.
3. Runtime executes — The actual function or API call is made outside the LLM.
4. Result returned — The output is injected back into the model's context.
5. Agent continues — The model decides whether to call another tool or produce a final answer.

Common Types of Tools

• Web search — Find current information from the internet.
• Web scraping — Extract structured content from a specific URL (e.g., KnowledgeSDK's /v1/scrape).
• Knowledge extraction — Parse a page into structured fields like product names, prices, and descriptions (KnowledgeSDK's /v1/extract).
• Code execution — Run Python or JavaScript and capture the output.
• Database queries — Read or write records from a SQL or NoSQL store.
• File operations — Read, write, or transform files.
• Email and calendar — Send messages or schedule events.

Why Tool Use Is Transformative

Tool use is what elevates LLMs from question-answering systems into autonomous agents. Consider a competitive intelligence agent:

• Without tools, it can only recall what competitors it knows from training.
• With tools, it scrapes competitor websites in real time, extracts product details, and synthesizes up-to-date intelligence.

KnowledgeSDK is purpose-built to serve as the web intelligence layer for tool-using agents. Its /v1/extract endpoint transforms any URL into structured knowledge, and /v1/search enables semantic retrieval over previously extracted content — making it straightforward to equip an agent with robust research capabilities.

Best Practices

• Keep tool descriptions concise and precise — The LLM uses them to decide when and how to call a tool.
• Validate inputs before execution — Do not blindly execute whatever the model generates.
• Return structured outputs — Well-formatted tool results are easier for the model to reason about.
• Limit tool scope — Each tool should do one thing well; composability comes from combining focused tools.