TLDR: Anthropic has introduced a new ‘Code Execution with MCP’ approach, transforming how AI agents interact with external tools. This method allows agents to write and execute code to utilize Model Context Protocol (MCP) servers, drastically reducing token consumption by up to 98.7% and significantly improving efficiency, speed, and data privacy for complex AI workflows.
Anthropic has unveiled a significant advancement in the operational efficiency of AI agents with its new ‘Code Execution with Model Context Protocol (MCP)’ approach. This innovation addresses a critical scaling problem faced by traditional MCP agents, where extensive tool definitions and intermediate results consume excessive tokens, leading to high costs, increased latency, and context window limitations.
Traditionally, AI agents using MCP would load numerous tool definitions and stream all intermediate results directly through the model’s context. For instance, a workflow involving fetching a long sales meeting transcript and then updating a Salesforce record would pass the entire transcript through the model twice, adding tens of thousands of unnecessary tokens. This method becomes unsustainable when agents need to interact with many MCP servers and tools.
Anthropic’s solution restructures this pipeline by presenting MCP servers as code-level APIs within a code execution loop. Instead of directly calling tools, the AI model is now tasked with writing and running TypeScript code to import and compose these modules. The MCP client exposes each server as a set of code modules in a filesystem, allowing the agent to dynamically load only the necessary tools and process data within a sandboxed execution environment before returning only the final, relevant results to the model.
This paradigm shift yields remarkable quantitative improvements. Anthropic reports a concrete example where a workflow that previously consumed approximately 150,000 tokens was re-implemented using the code execution pattern. The new approach reduced token usage to about 2,000 tokens, representing a staggering 98.7% reduction. This substantial decrease in token consumption directly translates to lower operational costs and significantly improved latency. Furthermore, this method can increase task completion speed by up to 10 times.
Beyond efficiency, the ‘Code Execution with MCP’ approach offers enhanced data privacy and security. Sensitive data can be preprocessed within the secure execution environment, ensuring that only filtered or aggregated information is passed back to the model. It also improves the composability and maintainability of tools, allowing for more powerful and context-efficient control flow, such as executing complex conditional logic in a single step rather than multiple agent-model interactions.
While this engineering optimization doesn’t alter the core capabilities of AI agents, it significantly enhances their practical deployment and scalability. However, it also introduces new considerations, such as the need for robust security in code execution environments and the potential for incorrect agent-generated code to cause task failures or ‘hallucination’ problems. Commentators also note that this approach might shift some computational costs to the user’s local environment and raises questions about ensuring the reliability of the tools for which the agent writes code.
Also Read:
- GitHub Unveils AgentHQ at Universe 2025, Revolutionizing Developer Workflows with AI Agents
- Google Enhances AI Agent Development with New Go Support in Agent Development Kit
Overall, Anthropic’s ‘Code Execution with MCP’ is a pivotal step towards building more efficient, cost-effective, and scalable AI agent systems, applying established software engineering patterns to address the unique challenges of large-scale AI deployments.
