Launches
Smaller tool outputs, bigger bills? New agent benchmark tests the whole session
Community reports 4 sources
B
Beyond Compression Ratios
Project Oct 17, 2026
A community benchmark dataset comparing total agent-session prompt tokens and answer correctness with raw, compressed and truncated MCP tool outputs.
Visit primary link
2 sources attached
What happened lately
Beyond Compression Ratios timeline
Beyond Compression Ratios publishes logged agent episodes, task specifications and output measurements for studying context compression. Its English card describes comparisons with local open-weight models through Ollama, including real MCP git, filesystem and SQLite servers over the Flask repository and Chinook sample database.
The dataset is licensed CC BY 4.0 and retains attribution requirements for its source data. Its results describe the author’s recorded small-model, short-session experiments rather than a universal ranking of compression systems. The card documents a synthetic-run environment-variable bug, timeouts and limits on reproducing synthetic worlds. Real-server tasks use public servers and data.