Multi-Function Patterns

Modern AI models can call multiple functions in a single turn, enabling efficient orchestration.

Parallel Function Calls

When the model identifies independent operations, it can request them simultaneously:

// Model response might include multiple tool calls:
// 1. get_weather("Tokyo")
// 2. get_weather("London")
// 3. get_exchange_rate("JPY", "GBP")

// Execute all in parallel
const results = await Promise.all(
  toolCalls.map(call => executeFunction(call.name, call.arguments))
);

Sequential Chains

Some operations depend on previous results:

User: "Book a flight from NYC to the cheapest warm destination"
→ Call 1: search_warm_destinations() → ["Cancún", "Miami", "Bahamas"]
→ Call 2: search_flights("NYC", "Cancún") → $350
→ Call 3: book_flight("NYC", "Cancún", $350) → Confirmation

Provider Differences

Optimization Tips

• Group independent calls for parallel execution
• Use tool_choice to force specific function selection when needed
• Limit the number of tools to reduce decision complexity
• Consider splitting large tool sets across specialized agents

Common Pitfall: Providing Too Many Tools at Once

Models perform significantly worse when given more than 10–15 tools simultaneously. The tool selection accuracy drops as the model must reason over a larger decision space. Group tools by task domain and only provide the relevant subset for each user intent:

const toolSets = {
  shopping: [searchProductsTool, addToCartTool, checkoutTool],
  support: [lookupOrderTool, createTicketTool, processRefundTool],
  account: [updateProfileTool, changePasswordTool, viewHistoryTool],
};

// Classify the user's intent first, then provide the relevant tool set
const intent = await classifyIntent(userMessage);
const tools = toolSets[intent] ?? toolSets.support;

Best Practice: Log All Tool Calls for Auditing

Every tool call should be logged with its input, output, and result. This creates an audit trail for compliance and an invaluable debugging dataset:

async function executeWithAudit(toolName: string, args: unknown) {
  const logEntry = { toolName, args, timestamp: new Date(), requestId };
  try {
    const result = await executeFunction(toolName, args);
    await db.toolCallLogs.create({ ...logEntry, status: "success", result });
    return result;
  } catch (error) {
    await db.toolCallLogs.create({ ...logEntry, status: "error", error: String(error) });
    throw error;
  }
}

Named Product Reference: OpenAI Assistants API with Tools

OpenAI's Assistants API manages the tool call loop automatically — you define tools and provide a run, and the API handles the back-and-forth between model responses and tool execution. It also supports code interpretation (Python sandbox) and file search as built-in tools. For developers who want function calling without implementing the agent loop manually, the Assistants API is a significant time-saver.