Anthropic slashes MCP token bloat with code

This feels like the point where MCP stops being a better tool-calling interface and becomes a runtime architecture story. The model should plan and write code; the sandbox should handle the repetitive, stateful work.

• –Progressive disclosure keeps tool catalogs proportional to the task, not the total number of integrations.
• –Filtering and summarizing data locally means the model sees fewer tokens and less raw customer data.
• –Moving loops, retries, and conditionals into code should make multi-step agent workflows far less brittle.
• –The pattern is model-agnostic, so teams can adopt it without betting on a single vendor stack.
• –The real tradeoff shifts to sandbox security, permissions, and observability, which now matter as much as prompt quality.