Understand this story
On November 17, 2026, UC Berkeley, Microsoft Research, and Microsoft Security released the open-source plugin TokenEconomizer. The plugin learns from coding agents' historical sessions in repositories, reducing the token cost of subsequent sessions without modifying the agents or the underlying models. Across 277 SWE-Bench tasks and five agent-model combinations, average cost dropped by 18.4–30.9%; all 30 repository-combination pairings saved money and task success rates remained broadly stable; Web-Bench sequential development tasks cut costs by 16.0–21.2%, and security audit tasks by 6.4–19.6%. Its mechanisms include micro-skills that record repository-specific knowledge, pruning infrequently used tools, and having a judge model step in to suggest a change of approach when an agent gets repeatedly stuck.
AI-generated from public reports
Developments
1 confirmed developments
Microsoft Research and others release open-source plugin TokenEconomizer
Source:Microsoft Research
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 17
Microsoft ResearchFirst-partySelected
Microsoft Research and others propose TokenEconomizer: enabling coding agents to learn from historical sessions to reduce token costs
Microsoft Research, in collaboration with UC Berkeley and Microsoft Security, proposes an open-source plugin called TokenEconomizer, which learns from coding agents' historical sessions in repositories to reduce the token overhead of subsequent sessions without modifying the agents or the underlying models. Across 277 SWE-Bench tasks and five agent-model combinations, the average cost is reduced by 18.4–30.9%; all 30 repository-combination pairings save money, and task success rates remain generally stable. On Web-Bench sequential development tasks, costs are reduced by 16.0–21.2%, and on security audit tasks by 6.4–19.6%. The mechanisms include recording micro-skills of repository-specific knowledge, pruning infrequently used tools, and having a judge model step in to suggest a change of approach when an agent gets repeatedly stuck.