We built something similar an A/B testing framework that measures Unblocked's impact on real AI coding agents.
It spawns agent CLIs (Claude Code, Codex, Cursor, GitHub Copilot) with and without Unblocked's MCP server attached, then statistically compares the results: https://github.com/unblocked/unblocked-harness-compare
We likewise measured token savings, (wall clock) time, # tool calls, and # turns.