Building Tools for AI Agents

Building Tools for AI Agents

Workflow

1. Define Purpose

   • Identify what agents need to accomplish with this tool
   • Determine if existing tools can be consolidated
   • Plan the tool's interface for agent consumption

2. Design Interface

   • Choose descriptive, namespaced tool names
   • Define parameters with clear types and descriptions
   • Design output format for maximum signal, minimum tokens

3. Implement

   • Build with token efficiency in mind
   • Add pagination, filtering, sensible defaults
   • Return semantic identifiers, not raw IDs

4. Validate

   • Test with real agent workflows
   • Check token consumption patterns
   • Verify error messages guide agents toward solutions

Design Principles

Tool Consolidation

• More tools don't lead to better outcomes
• Combine related operations into single tools
• Example: schedule_event that checks availability AND creates event
• Avoid simple CRUD-style tools that require multiple calls

Namespacing

• Prefix related tools with service name: asana_projects_search, asana_users_search
• Group by domain to help agents distinguish functionality
• Use consistent naming patterns across tool families

Meaningful Context

• Return high-signal information, not raw data dumps
• Resolve cryptic UUIDs to human-readable identifiers
• Include response_format parameter (concise/​detailed) for flexibility
• Surface relevant metadata agents need for next steps

Token Efficiency

• Implement pagination with sensible defaults
• Add filtering parameters to reduce unnecessary data
• Truncate large responses intelligently
• Prefer structured output over verbose prose

Tool Descriptions

• Invest heavily in clear, explicit descriptions
• Describe what the tool does, when to use it, and what it returns
• Include parameter constraints and valid values
• Small description improvements yield large performance gains

Anti-Patterns

• Creating many granular tools instead of consolidated operations
• Returning raw IDs that agents can't interpret
• Omitting pagination on potentially large result sets
• Vague tool descriptions that leave agents guessing
• Error messages that don't help agents recover
• Requiring agents to make multiple calls for common workflows

MCP-Specific Patterns

Tool Registration

• Use descriptive name and description in tool schema
• Define inputSchema with JSON Schema for parameters
• Mark required vs optional parameters explicitly

Response Format

• Return structured JSON for predictable parsing
• Include success/​error indicators
• Provide actionable error messages