Agent Tool Builder

Most agent failures trace back to poorly described tools. The LLM never sees your implementation, just the schema and description, so a vague "gets stock price" fails while "retrieves current trade price in USD for valid NYSE/NASDAQ ticker symbols, not historical data" works. This walks through schema design, parameter descriptions, enum constraints, and error handling that actually helps Claude recover. The key insight is counterintuitive: you need fewer than 20 tools total or the agent gets confused, and every error must be instructive, never silent. Includes the emerging MCP standard and Anthropic's beta input examples feature, which apparently bumps accuracy from 72% to 90% on complex operations by showing concrete patterns schemas can't express.