Cursor changed its agent harness to reduce user token costs by 7% without lowering agent quality, the company says.
Cursor’s breakdown attributes the savings to a system prompt trimmed by roughly 66%, dynamic tool loading that cut static-context description tokens by 60%, and cache changes that reduced cold misses by 20%. It also reports a 1.6% drop in cache-read tokens after numbering only every tenth line. These measurements are vendor-reported and have not been broadly validated.
Instead of packing every tool definition into persistent context, Cursor now keeps high-frequency tools loaded and fetches the rest on demand. It also separates stable setup from variable input through a “phantom user message,” preserving more reusable context.
For agent builders, the pattern is practical: prune persistent instructions, lazy-load rarely used tools, and isolate changing context before tuning the model itself.