How Anthropic Stole a Page from IDEs to Save AI Agents from Themselves

Anthropic AI optimization techniques

Anthropic’s AI agents are shrinking their memory footprint while growing their capabilities—by stealing a trick from human software engineers.

Anthropic’s Model Context Protocol (MCP) faced a "startup tax" where preloading all tool definitions consumed 67,000+ tokens per setup, sacrificing 33% of Claude Code’s 200,000-token limit.

Thariq Shihipar, Anthropic technical staff, explained: "We've found that MCP servers may have up to 50+ tools."

"The token savings are dramatic: from ~134k to ~5k in Anthropic’s internal testing,"

Internal benchmarks showed Opus 4 accuracy improved from 49% to 74%, and Opus 4.5 from 79.5% to 88.1% with Tool Search.

The update shifts from brute-force loading to "lazy loading," mirroring modern IDE practices like VSCode’s non-preloaded extensions. Product availability: Free for Claude Code users; developers need to implement ToolSearchTool for MCP clients.