TokenEconomizer reduces coding agent costs using repository-specific session experience

Key details TokenEconomizer is an open-source plugin that learns repository-specific knowledge and tool-use patterns from completed session logs. It applies brief knowledge notes, removes unnecessary tool descriptions, and intervenes in repeated failures without changing the existing agent or model. Across 277 SWE-Bench tasks from six repositories and five agent-model combinations, it reported average cost reductions of 18.4–30.9%, with success rates generally maintained. Use cases As tasks are performed sequentially in the same repository, test-running methods and repeated failures from earlier sessions inform subsequent sessions. Notes are added after patterns are confirmed across multiple sessions and discarded if later evidence contradicts them. Costs and limitations The plugin itself requires no API key, and learned information is stored on the developer's device. The cost-saving figures include the controller's own model-call costs, but costs may increase on initial tasks. Licensing and underlying agent fees are unconfirmed.

Visit original source