Code Execution with MCP: Building More Efficient Agents

engineering blog · source date 2025-11-04 · added 2026-08-07 15:52:43 · updated 2026-08-07 15:52:43 · Open original blog

Loading every MCP tool definition into context burns tokens before the agent does anything.
Intermediate tool results flow through the model even when the model only needs a filtered summary.
Tool selection degrades as the tool count grows.

Presents MCP servers as code APIs the agent explores on a filesystem and calls via generated code, rather than as a flat list of tool schemas.
Progressive disclosure loads only the tool definitions actually needed.
Intermediate data is filtered in-sandbox before it ever reaches the model.
Reports cutting one Drive-to-Salesforce workflow from 150k to 2k tokens (98.7%).

"Code Mode" is one of the most-discussed agent-tooling patterns of the period, reframing tool selection as code plus progressive disclosure.
It is cross-referenced by Cloudflare's Code Mode, later Anthropic advanced-tool-use releases, and much of the skill/tool-learning literature (semantic tool discovery, lazy schema loading).
Caveat: out-of-window (November 2025). The 98.7% figure is a single-workflow, Anthropic-reported result and savings vary widely by task. Code-as-tools shifts complexity into sandboxing and security rather than eliminating it.

Comments

No comments yet.