Code execution with MCP: building more efficient AI agents

Code execution with the Model Context Protocol (MCP) enables AI agents to handle thousands of tools more efficiently by loading tool definitions on-demand rather than upfront, reducing token consumption by up to 98.7%. Instead of passing all tool definitions and intermediate results through the model's context window, agents can write code to interact with MCP servers, filtering and transforming data in the execution environment. This approach provides progressive tool discovery, context-efficient data processing, privacy-preserving operations through tokenization, and state persistence across operations. While adding infrastructure complexity, code execution applies established software engineering patterns to solve agent scalability challenges.