Claude Advanced Tool Use (www.anthropic.com)

Anthropic announced three features to make Claude scale across hundreds or thousands of tools: a Tool Search Tool for on‑demand discovery, Programmatic Tool Calling (PTC) so Claude can orchestrate tools via code in a sandboxed execution environment, and Tool Use Examples to teach usage patterns beyond structural JSON schemas. The combo solves two core problems for real-world agents: exploding token costs from loading large MCP toolsets and “context pollution” from intermediate tool outputs, while reducing repeated inference calls and improving accuracy on complex workflows (e.g., Claude for Excel can modify huge sheets without overloading context). Technically, tools can be marked defer_loading: true and kept out of context until the Tool Search Tool (regex, BM25, or custom) returns relevant matches, cutting token overhead from ~77K to ~8.7K in an example (≈85% reduction) and boosting internal MCP accuracy (Opus 4: 49→74%; Opus 4.5: 79.5→88.1%). PTC lets Claude emit Python orchestration that runs in a Code Execution tool, fetching and processing data off‑context so only final results enter the model — yielding average token reductions (~43,588→27,297, 37% savings), fewer inference round‑trips, lower latency and measurable benchmark gains (knowledge retrieval 25.6→28.5%; GIA 46.5→51.2%). Together these features enable scalable, accurate agent orchestration across large tool libraries and messy production data.