Microsoft's TokenEconomizer Cuts Coding Agent Costs 18–31% on SWE-Bench
unixterminal · x · 2026-10-30
Microsoft Research and UC Berkeley open-sourced TokenEconomizer (500+ stars), an experience-driven black-box cost control tool for coding agents.
- Plugs into Claude Code, Codex, Copilot CLI, and Pi; learns from past sessions in a repo — remembers commands that work there, drops unused tools, and nudges the agent out of failure loops.
- Measured on SWE-Bench tasks: 18.4–30.9% lower mean cost across five agent–model combinations, cheaper in all 30 repository × combination cells, with Pass@1 up in 4 of 5.
- Author list includes Ion Stoica, Bin Yu, and Jianfeng Gao.
More from coding & agent
- Measuring compaction on a 400k-token session: 38s and $1.99 per run — brandon_galang · 2026-10-30
- pstack's /statechart skill ships with a mini TLC compiler that helped find real distributed-systems bugs — DavidKPiano · 2026-10-30
- First video demo: using Grok bot for daily workflow, outer vs inner loop — fbaptista · 2026-10-30
- DaVinci Resolve 21.1.2 expands scripting SDK for agentic video editing — dkundel · 2026-10-30
- 'People thought RAG would replace structured databases' — AI circle revisits LangChain hype — teodorio · 2026-10-30
- Tokyo Agent Hackathon: 40+ builders burn 580M tokens in 5 hours, ship 20+ projects — DavidBennett__ · 2026-10-30