Anthropic killed Tool calling

2026-02-22 · Source: AI Jason · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, long

Summary

Anthropic has released significant updates to its agent tool calling capabilities, which are crucial for building agents that handle complex, long-running tasks. These updates, termed "tool calling 2.0," address limitations of traditional tool calling, such as inefficiency and excessive token consumption, which arise from the model's reliance on generating JSON parameters for each function call. The core improvements include programmatic tool calling, which allows models to generate code for multi-function workflows, reducing round trips and token usage by 30-50%. Dynamic filtering for the web fetch tool reduces token consumption by an average of 24% by extracting only relevant content from HTML. A new tool search concept optimizes context windows by dynamically retrieving relevant tool schemas, saving up to 80% of tokens for agents with many tools. Finally, tool use examples improve accuracy in handling complex tool parameters from 72% to 90% by providing explicit usage examples.

Key takeaway

For AI Engineers building agents for complex, multi-step tasks, Anthropic's new tool calling features offer substantial efficiency and accuracy gains. You should integrate programmatic tool calling to reduce token consumption by 30-50% and improve deterministic workflows. Additionally, leverage dynamic filtering for web data and implement tool search for agents with over 10 tools to optimize context window usage and enhance overall agent performance.

Key insights

Anthropic's tool calling updates enhance agent efficiency and accuracy for complex tasks by optimizing token use and improving tool interaction.

Principles

• Code generation is more efficient for tool orchestration than JSON parameter generation.
• Context window optimization is critical for LLM performance, even with large models.
• Explicit examples improve LLM accuracy in complex tool usage.

Method

Programmatic tool calling enables LLMs to write code for multi-function workflows, using loops and conditionals to manage tool execution and data passing, thereby reducing token consumption and improving determinism.

In practice

• Implement programmatic tool calling by enabling a code execution function and specifying "allowed callers" for tools.
• Use dynamic filtering with the web fetch tool (version 2026209) to reduce token consumption for web content.
• Configure "deferred loading" for tools to enable dynamic tool search, optimizing context for agents with many tools.

Topics

• AI Agent Tooling
• Programmatic Tool Calling
• LLM Context Optimization
• Dynamic Function Calling
• Tool Use Examples

Best for: Machine Learning Engineer, AI Engineer, Software Engineer

Related on AIssential

• Anthropic Launches "Code Review" to Fix AI Code Security Issues — Anthropic's new AI code review tool enhances software quality by automating the detection of bugs and security risks in AI-generated code.
• To Call or Not to Call: A Framework to Assess and Optimize LLM Tool Calling — LLM tool-calling decisions, especially for web search, often misalign with true necessity and utility.
• Implement Anthropic's Context Engineering Framework with open source models — Dynamic context engineering mitigates "context rot" in LLM agents by managing information flow and computational load.
• Anthropic Steals OpenAI's Crown — The AI landscape is rapidly diversifying with advanced models, specialized tools, and integrated agent platforms.
• A New Labour Model: Salesforce to Spend US$300m on Anthropic — Salesforce's US$300M investment in Anthropic tokens signals a shift to AI-driven "new labor models" for productivity.
• Anthropic Introduces Managed Agents to Simplify AI Agent Deployment — Anthropic's Managed Agents simplify AI agent deployment by abstracting infrastructure complexities into a managed execution layer.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Jason.