Introducing Advanced Tool Use in the Claude Developer Platform

AI agents now need to work across hundreds or thousands of tools, but putting every definition into the prompt at once has clear limits. Tool definitions and results eat up 50,000+ tokens, and problems like picking similarly named tools by mistake or entering wrong parameters occur frequently.

Anthropic unveiled 3 features to solve these problems.

• Tool Search Tool: Searches for and loads only the tools that are needed, so the entire tool library doesn't have to be loaded into the context window all at once.
• Programmatic Tool Calling: Lets Claude call tools within a code execution environment, reducing context pollution and inference costs from repeated natural-language-based calls.
• Tool Use Examples: Provides a standard for teaching actual usage patterns through examples, which a JSON schema alone can't reveal.

Tool Search Tool defers loading of tools with defer_loading: true, and expands the relevant tool definitions only when a search is needed. For example, in a multi-server environment combining GitHub, Slack, Sentry, Grafana, and Splunk, about 55K tokens are used just for tool definitions, but with this feature, initial loading drops to about 500 tokens, with only the needed 3-5 tools loaded additionally. Anthropic explains that this results in an 85% reduction in token usage and 95% preservation of the context window, and in MCP evaluations, accuracy improved from 49%→74% for Opus 4 and from 79.5%→88.1% for Opus 4.5.

This approach is also advantageous for prompt caching. Since deferred tools aren't included in the initial prompt, the system prompt and core tool definitions remain cacheable as-is. However, since a search step is added, this is especially effective when there are many tools and the context savings outweigh the search cost.

Programmatic Tool Calling coordinates repetitive tool calls through code rather than natural-language back-and-forth. For example, when checking whether travel expenses have exceeded budget, Claude fetches team member lists, budgets, and expense records in parallel via Python code, compares them, and leaves only the final result in the context. Since the large intermediate data doesn't accumulate in the model, it allows for more stable control even in complex workflows.

Results were also presented. In complex research tasks, average token usage dropped from 43,588→27,297, a 37% reduction, internal knowledge retrieval improved from 25.6%→28.5%, and the GIA benchmark improved from 46.5%→51.2%. In other words, the more a task involves many tool calls and large intermediate results, the greater the benefit of code-based orchestration.

The common goal of these features is clear: load tool definitions only when needed, control tool calls through code, and convey correct usage through examples — simultaneously boosting both the scalability and accuracy of large-scale agent systems.