This open-source plugin from Microsoft Research and UC Berkeley reduces token usage in AI coding agents by 18.4–30.9%. It learns from completed sessions to prune unused tools, inject repository-specific notes, and break failure loops. The system runs alongside existing agents without modifying the underlying model or requiring new API keys.
Savings depend on working in the same repository repeatedly, as the tool needs prior session history to be effective. It is not a solution for one-off tasks or new repositories, though it showed modest gains in security auditing. Use it if you work on long-lived codebases and want to reduce recurring overhead.
my other side: ivo.blogspot.com
Leave a Reply