I optimized Cursor's agent harness. Check it out for nitty gritty and interesting details on where token spend comes from.
This one was actually a tricky rollout because tokens are so important to our top line. But we care about keeping low prices thanks to efficiency.
We've reduced token costs in Cursor by 7% with no drop in agent quality.
Savings came from tighter prompts, selective tool loading, better caching, and compressed file reads.