Anthropic’s latest update to the Claude Developer Platform marks another turning point for developers aiming to build smarter, more scalable AI-powered systems.
As tool libraries expand, traditional methods of loading every tool definition into an agent’s context window are hitting limits. Anthropic solves this with three new features designed to unlock flexibility and performance:
- Tool Search Tool: Agents dynamically discover and load only the tools they need, keeping context windows lean.
- Programmatic Tool Calling: Multi-step workflows run via orchestrated code, slashing token usage and inference time.
- Tool Use Examples: Practical usage samples teach agents to handle complex inputs more accurately than definitions alone.
In practice, these features streamline processes that were previously cumbersome or impossible, delivering measurable improvements in workflow efficiency and correctness.
Tool Search Tool: Discovery Without Context Overload
Large toolsets can quickly exhaust an agent’s context window, leading to errors and inefficiency. The Tool Search Tool allows Claude to defer loading tool details, accessing them only as needed through optimized search strategies. This preserves up to 95% of the context window and improves tool selection accuracy, internal benchmarks saw Opus 4 accuracy rise from 49% to 74%, and Opus 4.5 from 79.5% to 88.1%. For organizations with expansive tool libraries, this innovation is a game-changer.
Programmatic Tool Calling: Orchestrate Smarter, Not Harder
Rather than cluttering context with every intermediate result, Programmatic Tool Calling lets developers orchestrate workflows using code. Claude receives only the final, relevant results freeing up context and boosting performance, especially for large datasets or multi-step tasks. For example, Anthropic showed how evaluating expense compliance across thousands of records can now be completed with a single, efficient script, cutting token usage by 37% and reducing latency.
Tool Use Examples: Clarity Through Demonstration
Schema definitions outline structure but they don’t always communicate real-world usage. Tool Use Examples solve this by providing concrete sample calls directly alongside schemas. This approach clarifies ambiguous conventions and complex parameters, boosting accuracy in parameter handling from 72% to 90% in Anthropic’s tests. It’s especially helpful for tools with intricate or nested inputs where mistakes are common.
Best Practices: Layer and Tailor for Your Use Case
Anthropic recommends selecting the right combination of features for each workflow:
- Deploy Tool Search Tool for large-scale or context-sensitive operations.
- Use Programmatic Tool Calling when workflows demand orchestration or handle big data.
- Incorporate Tool Use Examples to reduce errors and clarify expectations.
Clear system prompts and thoughtful tool documentation further enhance performance. When orchestrating with code, document output formats and limit orchestration to compatible tools for best results.
Start Building With Advanced Tool Use
Developers can access these powerful features in beta by enabling them with specific headers on the Claude Developer Platform. Comprehensive documentation and sample code are available, making it easy to integrate dynamic tool discovery, programmatic workflows, and example-driven guidance into your systems.
Takeaway: Setting a New Standard for AI Agent Automation
With dynamic discovery, efficient workflow orchestration, and real-world usage guidance, Claude evolves from a basic function caller into a robust, context-aware orchestrator. These breakthroughs pave the way for a new generation of AI agents that are capable, reliable, and ready for real-world complexity.
Source: Anthropic Engineering Blog
How Anthropic’s Claude Platform Redefines Advanced Tool Use for AI Agents