Code Mode
Instead of invoking tools one call at a time, the agent writes code that calls MCP servers exposed as code APIs, so only the definitions and results it actually needs enter context.
Plain English
The agent writes a small script against your tools instead of trading messages with each one, keeping the middle of the workflow out of the model's view.
Example
With twenty MCP servers connected, the agent reads a typed github.createIssue stub, writes a loop that calls it fifty times, filters results in the sandbox, and only a one-line summary reaches the model.
Why it matters
Anthropic measured a 150,000-token workflow shrinking to 2,000 (98.7%) - the current best answer to tool definitions eating the context window.