Introducing advanced tool use on the Claude Developer PlatformNov 24, 2025

• AI agents work best when they can access tons of tools without loading all their instructions upfront—imagine an IDE helper that can tap into git, testing, and deployment tools only when needed, instead of memorizing every single one
• When agents have to describe what they're doing in natural language for every tiny action, it wastes a lot of tokens and piles up unnecessary info in memory; letting them write code instead (with loops, if-statements, etc.) is way more efficient for complex tasks
• Tools need real examples of how to use them, not just technical specs—a JSON schema tells you what's *allowed*, but it doesn't explain when to use optional features or what your specific API actually expects in practice
• Anthropic released three new features to fix these problems: a tool search function (so Claude can find the right tool from thousands without memorizing them all), the ability to call tools directly from code (cutting down on wasted tokens), and a standard way to show examples of proper tool usage
• In real testing, these features made previously impossible things possible—like Claude for Excel being able to work with massive spreadsheets without overwhelming the AI's memory limits