Skip to content
Microsoft Research· Aaron J. Liu, Weijia Xu, Chandan Singh, Jeevana Priya Inala, Bin Yu, Ion Stoica, Michel Galley, Jianfeng Gao·· 40 min agoPicksAI score66

TokenEconomizer: Microsoft Research plugin cuts AI coding agent costs 18.4–30.9% on SWE-Bench

TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient

AI summary

TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research and Microsoft Security, reduced costs by 18.4–30.9% across five AI coding agent and model combinations on 277 SWE-Bench tasks from six repositories, with task success rates broadly stable. It learns from an agent's past sessions in a repository through three mechanisms: micro-skills that add one-line repository notes, pruning of unused tools, and failure-loop intervention that suggests a different approach when a judge model finds the agent stuck. Savings also reached 16.0–21.2% on Web-Bench and 6.4–19.6% on security auditing, and cumulative costs fell below the default in every repository by the sixth task.

Why it matters

A Microsoft Research and UC Berkeley plugin cuts coding-agent token costs by 18.4–30.9% on SWE-Bench by learning from a repository's past sessions without touching the model.

Source:Microsoft Research · microsoft.com