Anthropic Just Killed Tool Calling

Entropic's New Developer Tools: Programmatic Tool Calling

Introduction to Programmatic Tool Calling

• Entropic has introduced significant developer tools with the release of set 46, focusing on programmatic tool calling.

• This feature allows agents to make specific calls to tools via code instead of loading everything into the context window, saving tokens and improving accuracy.

Advantages Over Traditional Methods

• The engineering work by Anthropic highlights their commitment to innovation in tool calling, which is more effective than traditional JSON structures.

• LLMs (Large Language Models) are better suited for code execution rather than conventional tool calling due to their training background.

Industry Impact and Adoption Trends

• Anthropic's innovations often lead to industry-wide adoption, as seen with MCPS (Model Context Pro Protocol).

• The context window problem is exacerbated by protocols like MCP, leading to inefficient use of space during user interactions.

Context Engineering and Its Importance

• Context engineering aims to optimize what information is loaded into the context window, discarding unnecessary data.

• Tool calls significantly contribute to context pollution; thus, optimizing them can enhance performance.

How Programmatic Tool Calling Works

• In programmatic tool calling, coding agents write code in a sandbox environment for invoking tools rather than making direct calls.

• This method reduces token usage since only final outputs are returned from the sandboxed process.

Timeline of Developments in Programmatic Tool Calling

• Cloudflare published a report in September 2025 advocating for programmatically invoking tools within an MCP server, showing potential token savings of 30%–80%.

• Anthropic echoed these findings in November 2025 with their article on building efficient agents using MCP.

Recent Advancements and Community Response

• Anthropic released advanced tools including a search function that optimizes token usage further.

• The open-source community rapidly adopted these concepts, leading to implementations across various platforms like Blocks Goose Agent and Light LLM.

Conclusion on Current State and Future Directions

• As of now, these advancements have moved beyond beta testing into full support with dynamic filtering capabilities for web searches.

5.2 API Enhancements and Tool Support

Key Insights on LLMs and Code Generation

• Version 5.2 has introduced support for over 20 different tools via their API, enhancing the capabilities of large language models (LLMs).

• LLMs are trained on billions of lines of code, particularly effective in generating and understanding code but lacking in synthetic JSON tool calling formats.

• Anthropic's Sonnet 46 release includes two new tools: web search and dynamic filtering, which improve how agents interact with data.

Improvements in Web Search Capabilities

• The new features allow the model to write and execute code during web searches, filtering results before they enter the context window to enhance accuracy.

• Initial tests showed an average improvement of 11% in performance metrics while reducing input tokens by 24%, indicating significant efficiency gains.

Benchmark Performance Analysis

• The browser comp benchmark assesses an agent's ability to navigate websites for hard-to-find information; Sonnet improved from 33% to 46%, while OPUS increased from 45% to 61%.

• In the deep search QA benchmark, which evaluates finding multiple correct answers through web searches, Sonnet saw an F1 score rise from 52% to 59%.

Token Cost Considerations

• Token costs can vary based on how much code is generated for filtering; Sonnet's price-weighted token decreased while OPUS's increased due to more extensive coding requirements.

• This indicates that a reduction in output tokens does not always correlate with lower token costs; careful consideration is needed when evaluating performance.

Utilizing New Features Effectively

• Users employing the search API need only enable data fetching; Anthropic will automatically optimize token usage by returning only relevant information.

• Additional tools have exited beta status, including code execution sandboxes and programmatic tool calling, along with detailed documentation provided for user guidance.

Implementation Structure

• To implement these tools effectively, users must define what each tool does alongside its input/output schema within a structured format.

• Instead of traditional function calls, models like Cloud will now generate code directly for executing specific tasks as part of standard industry practices.