Claude Releases Engineering Guide for Building Token-Efficient AI Agents With MCP.

Anthropic has published new technical guidance on building AI agents that can manage multiple tools while consuming fewer tokens, addressing a key challenge as developers scale their agent deployments.

The engineering blog post focuses on code execution capabilities within the Model Context Protocol (MCP), a framework Anthropic introduced to standardize how AI assistants connect to different data sources and tools. The guidance arrives as organizations like Cognizant deploy Claude to hundreds of thousands of employees, making efficiency improvements increasingly critical at scale.

As AI agents handle more complex workflows, they often need access to dozens of tools simultaneously. Traditional approaches load descriptions and schemas for every available tool into the context window, rapidly consuming tokens even before the agent begins its actual work.

This creates a practical ceiling on agent capabilities. Developers face a choice between limiting tool availability or accepting higher latency and costs from bloated context windows.

The new MCP-based approach changes this dynamic. Instead of front-loading all tool information, agents can dynamically execute code to discover and interact with tools only when needed. This shifts token usage from initialization to execution, where it directly contributes to task completion.

The Model Context Protocol enables agents to run Python code within a sandboxed environment. Rather than describing every available API endpoint upfront, developers can expose a smaller set of MCP servers that the agent queries programmatically.

When an agent needs a specific capability, it writes and executes code to inspect available tools, read their documentation, and invoke them with appropriate parameters. This just-in-time discovery model means only relevant tool information enters the context window.

For example, an agent building a data pipeline might start by querying which database connectors are available, then load only the schema for the selected connector, then execute the specific query needed. Each step uses tokens purposefully rather than defensively.

The engineering post provides code samples showing how to structure MCP servers for optimal agent interaction. Key recommendations include organizing tools into logical groupings, providing clear naming conventions that agents can parse programmatically, and implementing consistent error handling that agents can interpret and recover from.

Developers building with Cursor, Replit, or other AI-assisted coding environments can apply these patterns to their own agent architectures. The approach works alongside existing tool-use capabilities in Claude, including the extended thinking mode recently launched in Claude Opus 4.5.

Companies like Brex have already demonstrated how Claude's code execution capabilities can power internal tools. The MCP integration extends this to multi-tool workflows where efficiency directly impacts user experience.

This release reflects a broader shift in how AI companies think about agent architecture. Rather than maximizing context windows to accommodate every possible need, the focus moves toward intelligent resource management and dynamic capability discovery.

For small business owners and builders using AI tools, this means more responsive agents that can handle complex tasks without the latency penalties of oversized prompts. As platforms like Make and n8n integrate AI agents into their automation workflows, token efficiency becomes a practical concern affecting monthly bills and response times.

The full technical guide is available on the Anthropic engineering blog. Developers can experiment with MCP through Claude's API or in supported development environments that have adopted the protocol.