Tool use (also called function calling) is the mechanism that turns a language model into an agent. It lets the AI call external APIs, query databases, and interact with the real world.

How Function Calling Works

1. You define a set of "tools" — functions the AI can call, with descriptions and parameter schemas
2. The AI receives a user request and your tool definitions
3. The AI decides whether to call a tool and with what arguments
4. Your code executes the tool and returns the result to the AI
5. The AI uses the result to formulate its response

Example: Weather Tool

You define:

• Tool name: get_weather
• Description: "Get current weather for a city"
• Parameters: { city: string, units: "celsius" | "fahrenheit" }

User asks: "What's the weather like in Tokyo?"

AI decides to call: get_weather({ city: "Tokyo", units: "celsius" })

Your code calls a weather API and returns: { temp: 22, condition: "partly cloudy" }

AI responds: "It's 22°C and partly cloudy in Tokyo right now."

Common Tool Types

• Search tools — Web search, database queries, document retrieval
• Action tools — Send emails, create calendar events, update CRM records
• Computation tools — Run calculations, execute code, process data
• Integration tools — Connect to Slack, GitHub, Jira, Stripe, etc.

Multi-Tool Orchestration

Powerful agents use multiple tools in sequence:

User: "Find our top 3 customers by revenue this quarter and draft a thank-you email to each"

Agent thinks: I need to query the database, then compose emails

1. Calls query_database for top customers
2. Calls get_customer_details for each customer
3. Calls draft_email three times with personalized content
4. Presents drafts for human approval

Parallel Tool Calls

Modern APIs support parallel tool calling — the AI can request multiple tool executions simultaneously:

• "What's the weather in Tokyo, London, and New York?" → 3 parallel get_weather calls
• This reduces latency significantly for multi-step tasks

Security Considerations

• Principle of least privilege — Only give agents access to tools they need
• Input validation — Always validate tool parameters before executing
• Rate limiting — Prevent agents from making excessive API calls
• Confirmation gates — Require human approval for destructive actions (delete, send, purchase)
• Audit logging — Record every tool call for accountability