Cursor cut agent token costs 7% without hurting quality

On September 23, 2026, Jediah Katz, Connor O'Keefe, and Calvin Yee wrote that longer agent runs spend more on context. Cursor improved harness efficiency and reduced user token costs about 7% without reducing agent quality.

They trimmed roughly 66% of the system prompt as models needed less hand-holding. Dynamic tool loading extends an earlier MCP idea: tools-as-dynamic-context cut tokens about 46.9% in sessions that called MCP, and the same idea for built-in tools cut static-context description tokens about 60%, while high-frequency read, search, edit, and shell tools stay static. Explicit prompt-cache breakpoints, available with the OpenAI API since GPT-5.6, plus moving variable setup into a phantom user message, cut cold cache misses about 20%. Line numbers on file reads only every tenth line meant about 1.6% fewer cache-read tokens. Subagent prompting and model selection were tightened to avoid an unnecessary coordination tax.

Those harness choices sit next to the Chapter 1 tooling you use when shipping apps with AI-assisted workflows.

Read original: Improved token efficiency for longer agent runs — Cursor

Related in the book: Chapter 1