Tim Dettmers releases CliffCompaction: million-token agent sessions at 50% less cost
Tim_Dettmers · x · 2026-09-23
Tim Dettmers shipped CliffCompaction as the 2nd release of his open-source week: an agent context-compaction tool that enables million-token sessions and saves 50% on token costs, integrating with Claude Code and Codex.
- The design is intentionally unintuitive: it deliberately throws away context after two compactions, because context drift mainly happens when an agent reads a lossy summary, assumes it has everything, and keeps working.
- It manages a separate compaction state between you and any OpenAI/Anthropic API; install with pip install cliffcompaction, enable with cliff enable, and it's integrated into his autonomous research harness bitsandbytes code.
- It works much better with full thinking traces; DeepSeek v4.1 and MiMo v2.6 Flash compact exceptionally well, and for long-horizon medium-complexity work open-weight models beat closed ones that hide reasoning traces.
More from coding & agent
- Reading an invoice and moving money should not share one permission: agent auth principles — TechNadu · 2026-09-24
- Dev jokes about adding 'digest context when agitated' to every agents.md — nptacek · 2026-09-24
- Harness searches hundreds of distributed agent sessions in 17ms with Cmd-O — dee_hw · 2026-09-24
- Open-source MCP server gives agents page-change tracking with SHA-256 capture certificates — Einperegrin · 2026-09-24
- LangChain Leans Into 'Composable AI': Jev and System One Models in the Agent Harness — hwchase17 · 2026-09-24
- Google launches $100K Gemma 4 Developer Agent Competition on Kaggle — michael_galkin · 2026-09-24