AI Patterns Tool Use Patterns

Tool Use Patterns

Quick Guide: Tool use (function calling) lets LLMs invoke external functions. The universal pattern is: define tool schemas (JSON Schema for parameters) -> send tools + message to LLM -> detect tool_use in response -> execute locally -> return result to LLM -> repeat until the model responds with text. Guard every loop with a max-step limit, validate all tool inputs before execution, and return structured errors so the model can recover. Use tool choice control (auto, required, none, specific tool) to steer model behavior.

CRITICAL: Before Using This Skill

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST guard every tool loop with a maximum step limit -- unbounded loops risk infinite API calls and runaway costs)

(You MUST validate all tool input arguments before execution -- LLM-generated arguments are untrusted input)

(You MUST return structured error messages to the model when tool execution fails -- never silently swallow errors or return empty results)

(You MUST use JSON Schema for tool parameter definitions -- all major providers require this format)

(You MUST treat tool definitions as token cost -- every tool schema is sent on every API call, so keep descriptions concise but precise)

Auto-detection: tool use, function calling, tool_calls, tool_use, tool call loop, agent loop, tool definition, tool schema, toolChoice, tool_choice, parallel tool calls, human-in-the-loop, tool approval, agentic workflow, multi-step agent, tool result, tool error

When to use:

• Implementing LLM tool calling / function calling in any provider
• Building agent loops that call tools iteratively until a task is complete
• Handling parallel tool calls (multiple tools in one response)
• Reporting tool errors back to the model for recovery
• Controlling tool selection (auto, required, none, force specific)
• Adding human approval gates before dangerous tool execution
• Streaming responses that include tool calls

Key patterns covered:

• Tool definition schemas (JSON Schema for parameters, descriptions)
• The core tool call loop (send -> detect -> execute -> return -> re-send)
• Parallel tool calls (handling multiple calls in one response)
• Error handling (reporting tool failures back to the model)
• Tool choice control (auto, required, none, specific tool)
• Multi-step agent workflows with conversation state
• Human-in-the-loop approval patterns
• Type-safe tool definitions in TypeScript
• Security (input validation, sandboxing, least privilege)
• Streaming with tool calls

When NOT to use:

• Simple text generation without tool calling -- no tools needed
• Structured output / JSON extraction -- use your provider's structured output feature instead
• Provider-specific SDK patterns -- use your provider's SDK skill for SDK-specific APIs

Detailed Resources:

• examples/core.md -- Tool definitions, the tool call loop, error handling, type-safe tools
• examples/advanced.md -- Parallel tool calls, multi-step agents, human-in-the-loop, streaming, security
• reference.md -- Decision frameworks, provider comparison, anti-pattern checklist

Philosophy

Tool use is the mechanism that turns LLMs from text generators into agents. The model cannot execute code, query databases, or call APIs -- it can only request that your code does so by emitting structured tool calls. Your code is the executor; the model is the planner.

Core principles:

1. The model plans, you execute -- The LLM emits tool call requests with structured arguments. Your code validates, executes, and returns results. Never let the model execute arbitrary code directly.
2. Agents are loops -- Every agent, from a simple weather bot to a complex coding assistant, follows the same loop: LLM decides -> system executes -> results feed back -> repeat. Complexity comes from the tools and state, not the loop itself.
3. Tools are schemas -- A tool definition is a JSON Schema that tells the model what function exists, what parameters it takes, and when to use it. Better descriptions produce better tool selection and argument quality.
4. Errors are information -- When a tool fails, return a structured error message to the model. The model can often recover by retrying with different arguments, choosing a different tool, or explaining the failure to the user.
5. Defense in depth -- LLM-generated arguments are untrusted input. Validate schemas, enforce types, limit argument ranges, sandbox execution, and require approval for dangerous operations.

When to use tool calling:

• The task requires real-world data the model doesn't have (weather, database, APIs)
• The task requires side effects (sending email, creating records, file operations)
• The task requires multi-step reasoning with intermediate data lookups
• The task requires computation the model can't do reliably (math, code execution)

When NOT to use tool calling:

• The model can answer from its training data alone
• You only need structured JSON output (use structured output features instead)
• The "tool" is just prompt engineering disguised as a function
• You want deterministic behavior (tool calling adds non-determinism from model decisions)

Core Patterns

Pattern 1: Tool Definition Schema

Every tool definition has three parts: a name, a description, and a parameter schema. The description is the most important part -- it guides the model's decision to call the tool and how it constructs arguments.

// Provider-agnostic tool definition shape
interface ToolDefinition {
  name: string; // ^[a-zA-Z0-9_-]{1,64}$
  description: string; // When and why to use this tool
  parameters: JsonSchema; // JSON Schema for input arguments
}

Good Tool Definition

// Good: precise description, constrained parameters, .describe() on each property
const VALID_UNITS = ["celsius", "fahrenheit"] as const;

const getWeatherTool: ToolDefinition = {
  name: "get_current_weather",
  description:
    "Get the current weather conditions for a specific city. " +
    "Returns temperature, humidity, and conditions. " +
    "Use this when the user asks about current weather, NOT forecasts.",
  parameters: {
    type: "object",
    properties: {
      location: {
        type: "string",
        description: "City name and optional country code, e.g. 'London, UK'",
      },
      unit: {
        type: "string",
        enum: VALID_UNITS,
        description: "Temperature unit. Defaults to celsius if not specified.",
      },
    },
    required: ["location"],
  },
};

Why good: Description explains what the tool returns, when to use it, and when NOT to use it. Parameters have descriptions and constraints (enum). Named constant for valid values.

Bad Tool Definition

// Bad: vague description, no parameter descriptions, no constraints
const weatherTool: ToolDefinition = {
  name: "weather",
  description: "Gets weather", // BAD: too vague
  parameters: {
    type: "object",
    properties: {
      loc: { type: "string" }, // BAD: no description, cryptic name
      u: { type: "string" }, // BAD: no enum constraint
    },
  },
};

Why bad: Vague description causes incorrect tool selection, no parameter descriptions cause malformed arguments, no enum constraint causes invalid values, cryptic parameter names confuse the model

Pattern 2: The Core Tool Call Loop

The fundamental pattern for tool use. Send a message with tool definitions, check if the response contains tool calls, execute them, return results, and let the model generate a final response. See examples/core.md for the complete implementation.

const MAX_TOOL_STEPS = 10;

for (let step = 0; step < MAX_TOOL_STEPS; step++) {
  const response = await callLLM({ messages, tools });
  if (!response.toolCalls?.length) return response.text;
  messages.push({ role: "assistant", toolCalls: response.toolCalls });
  for (const tc of response.toolCalls) {
    messages.push({
      role: "tool",
      toolCallId: tc.id,
      content: JSON.stringify(await executeTool(tc)),
    });
  }
}

Key points: Always use a bounded for loop with a named constant (MAX_TOOL_STEPS) -- never while (true). Always append the assistant message (with tool calls) before tool results. Include graceful termination when the step limit is reached.

Pattern 3: Tool Execution with Error Handling

When a tool fails, return a structured error to the model instead of crashing. The model can often recover by retrying, choosing a different approach, or explaining the failure. See examples/core.md for the complete validation pipeline.

// Four-step validation: tool exists -> JSON parses -> schema validates -> execution succeeds
async function executeTool(toolCall: ToolCall): Promise<ToolResult> {
  const handler = toolRegistry[toolCall.name];
  if (!handler)
    return { success: false, error: `Unknown tool: "${toolCall.name}"` };
  const validated = handler.schema.safeParse(toolCall.arguments);
  if (!validated.success)
    return { success: false, error: `Invalid args: ${validated.error}` };
  try {
    return { success: true, data: await handler.execute(validated.data) };
  } catch (error) {
    return {
      success: false,
      error: `Tool failed: ${error instanceof Error ? error.message : "Unknown"}`,
    };
  }
}

Key points: Always validate tool existence, parse JSON arguments, validate against schema, and catch execution errors. Return structured { success, data?, error? } results so the model can recover. Never crash the loop on tool errors.

Pattern 4: Tool Choice Control

Control whether and how the model uses tools. All major providers support four modes.

type ToolChoice =
  | "auto" // Model decides whether to call tools (default)
  | "required" // Model MUST call at least one tool
  | "none" // Model MUST NOT call any tools
  | { name: string }; // Model MUST call this specific tool

When to Use Each Mode

// auto (default) -- let the model decide
const response = await callLLM({
  messages,
  tools,
  toolChoice: "auto",
});

// required -- force tool use (e.g., first step of an agent must act)
const response = await callLLM({
  messages,
  tools,
  toolChoice: "required",
});

// none -- disable tools (e.g., final response must be text only)
const response = await callLLM({
  messages,
  tools,
  toolChoice: "none",
});

// specific tool -- force a particular tool (e.g., classification step)
const response = await callLLM({
  messages,
  tools,
  toolChoice: { name: "classify_intent" },
});

Decision guidance:

When to use: When you need to constrain or guarantee tool behavior at specific points in a workflow

Pattern 5: Type-Safe Tool Definitions in TypeScript

Use a typed registry pattern with Zod schemas to get compile-time safety for tool definitions and runtime validation for execution. See examples/core.md for the complete ToolRegistry class implementation.

import { z } from "zod";

// Define tool with schema + execute function -- TypeScript infers argument types
function defineTool<T extends z.ZodType>(config: {
  name: string;
  description: string;
  schema: T;
  execute: (args: z.infer<T>) => Promise<unknown>;
}) {
  return config;
}

When to use: Any TypeScript project implementing tool calling -- Zod validates at runtime, TypeScript catches schema/handler mismatches at compile time

Decision Framework

Do You Need Tool Calling?

Does the task require information the model doesn't have?
+-- YES -> Tool calling (fetch data from APIs, databases, files)
+-- NO -> Does the task require side effects?
    +-- YES -> Tool calling (send email, create record, execute code)
    +-- NO -> Do you need structured JSON output?
        +-- YES -> Use structured output features (NOT tool calling)
        +-- NO -> Plain text generation, no tools needed

Which Loop Pattern?

How many tools might the model call?
+-- Single tool call per request
|   +-- Simple request-response with one tool execution
|   +-- No loop needed, just one round-trip
+-- Multiple sequential tool calls
|   +-- Use the bounded tool call loop (Pattern 2)
|   +-- Set MAX_TOOL_STEPS based on task complexity
+-- Multiple parallel tool calls in one response
|   +-- Execute all tool calls concurrently (Promise.all)
|   +-- Then return all results and loop
+-- Complex multi-step agent
    +-- Use the bounded loop with conversation state
    +-- Add human-in-the-loop for dangerous operations
    +-- Consider per-step tool filtering

How to Handle Tool Errors?

Tool execution failed. What to do?
+-- Return structured error to the model
|   +-- Include error message and context
|   +-- Model can retry, choose alternative, or explain failure
+-- NEVER: silently return empty result
+-- NEVER: crash the loop
+-- NEVER: retry automatically without telling the model
    (the model should decide whether to retry)

RED FLAGS

High Priority Issues:

• Unbounded tool loop (while (true)) without a step counter -- risks infinite API calls and runaway costs
• Executing tool arguments without validation -- LLM-generated arguments are untrusted input, treat like user input
• Swallowing tool errors silently (returning null or {}) -- the model cannot recover from failures it doesn't know about
• Allowing arbitrary code execution from tool arguments without sandboxing -- prompt injection can escalate to code execution
• Tool descriptions that say "Gets data" -- vague descriptions cause wrong tool selection and malformed arguments

Medium Priority Issues:

• Sending all tools on every API call when only a subset is relevant -- wastes tokens and confuses the model
• Not including the assistant message (with tool calls) in conversation history before tool results -- breaks the message sequence
• Using tool_choice: "required" without a fallback for when no tool makes sense -- forces meaningless tool calls
• Returning raw database rows or full API responses as tool results -- overwhelms context with irrelevant data; summarize or truncate
• Not logging tool calls and results -- impossible to debug agent behavior in production

Common Mistakes:

• Forgetting that tool call arguments arrive as a JSON string, not a parsed object -- always JSON.parse() before use
• Assuming the model will always call tools when tools are available -- with auto mode, it may respond with text directly
• Treating tool calling as structured output -- they solve different problems (actions vs data extraction)
• Putting business logic in tool descriptions instead of tool implementations -- descriptions guide selection, not execution

Gotchas & Edge Cases:

• Tool definitions consume tokens on every API call -- 10 tools with detailed schemas can use 1000+ tokens per request
• Parallel tool calls may arrive in any order -- never assume execution order matches definition order
• Some models hallucinate tool names or arguments that don't match any definition -- always validate the tool name exists in your registry
• Streaming responses with tool calls require accumulating partial JSON chunks before parsing -- the arguments arrive incrementally, not all at once
• Returning very large tool results (>4000 tokens) can push the conversation past context limits -- truncate or summarize large results
• The model maintains full conversation context including all tool calls and results -- long agent runs accumulate significant token usage
• Different providers use different message formats for tool results (role: "tool" vs content blocks) -- abstract this in your callLLM wrapper

CRITICAL REMINDERS

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST guard every tool loop with a maximum step limit -- unbounded loops risk infinite API calls and runaway costs)

(You MUST validate all tool input arguments before execution -- LLM-generated arguments are untrusted input)

(You MUST return structured error messages to the model when tool execution fails -- never silently swallow errors or return empty results)

(You MUST use JSON Schema for tool parameter definitions -- all major providers require this format)

(You MUST treat tool definitions as token cost -- every tool schema is sent on every API call, so keep descriptions concise but precise)

Failure to follow these rules will produce agents that run up API costs in infinite loops, execute unvalidated input, or silently fail without the model being able to recover.