The second practical experiment is out.
Classical MCP failed the same data-processing task in all 5 runs. With Code Mode, all 5 runs succeeded.
In the video, I look at what changed architecturally and why moving raw data processing out of the LLM context made such a difference.
Watch the video: