Tool Use Patterns

Design and implement tools that AI agents can reliably call in production applications.

Tool design principles

• Single responsibility — one tool does one thing well
• Clear naming — searchSkills not doSearch or skill_op
• Typed inputs — JSON Schema with required fields and descriptions
• Typed outputs — consistent return shape for each tool
• Idempotent when possible — safe to retry on failure

Function calling flow

1. Model receives the system prompt + available tool definitions
2. Model decides which tool to call (or none)
3. System validates the call, executes the tool, returns the result
4. Model processes the result and either calls another tool or responds

Error recovery

• Return structured error objects, not raw exceptions
• Include actionable guidance in error messages
• Implement automatic retry with exponential backoff
• Fallback to alternative tools when primary fails
• Log all tool errors for debugging

Composable tool chains

Build complex operations from simple tools:

• searchWeb → extractContent → summarize → formatResponse