Benchmarks that hold up
I design the experiment, write the harness, and publish the code, so engineers can rerun every number.
- agent-memory-bench: 20M vs 4.4M tokens for the same 200-message conversation, 4.5x fewer
- mcp-token-benchmark: MCP servers from Notion, GitHub, Slack, and Playwright measured, a 25x spread, and a 95% lighter redesign of Notion's, 17,161 down to 773 tokens
- screen-test: open benchmark for generative video models, judged by working filmmakers and engineers








