Cursor cut agent harness token usage by 7%

Cursor engineers described an agent harness optimization that reduced token costs by 7% without losing agent quality: the system prompt was shortened by about 66%, static tool descriptions by 60% through dynamic loading, and explicit cache breakpoints in the OpenAI API reduced cold cache misses by 20%.

Compiled by Sergey KostenchukPublished 2026-09-24Updated 2026-09-24

2026-09-24 Coding OpenAI

💻 Cursor cut agent token usage by 7%

Cursor engineers optimized the agent harness: the system prompt was compressed by 66%, static tool descriptions by 60% through dynamic loading, and cache breakpoints in the OpenAI API (GPT-5.6) reduced cold cache misses by 20%. Quality was not affected.

🌍 Cursor showed the mechanism for saving: change what the harness sends to the model, rather than asking the model to be frugal — rare tools are unloaded from the static context, and the cache is placed after stable layers. This directly improves the unit economics of agent products.

👤 Building your own harness around an LLM: measure costs by billing type, clean up protective instructions for older models, and keep only read, search, edit, and shell static. The prompt from Telegraph can be handed to an agent for refactoring.

Details: https://telegra.ph/Harness-token-efficiency-by-Cursor-09-24 Source 1: https://cursor.com/blog/improved-token-efficiency

Sources