Cursor cuts agent token costs 7%
Cursor says a series of context-assembly changes to its agent harness cut user token costs by 7% without degrading agent quality, as longer-running agents shift where spend accumulates.
Cursor said changes across its agent harness — the layer that assembles each request, decides what context is reused, and when work is split across agents — cut token costs 7% with no drop in agent quality.
The largest single change was to static context. Cursor trimmed roughly 66% of its system prompt, saying newer models no longer need explicit "DO NOT" and "You must" scaffolding or guards against long hash dumps, binary output, and emojis. Tool definitions, which grew through the year with additions like background shell monitoring and cloud subagents, were also pruned: most built-in tools are needed in fewer than 20% of conversations, so they now load dynamically. Cursor had applied the same approach to MCP tools earlier this year, reducing total tokens 46.9% across sessions that called an MCP tool. Reading, searching, editing and shell tools stayed in static context, along with ask_question — which some models hallucinated calls for — and create_plan for Plan Mode.
On caching, Cursor said it now uses the explicit cache breakpoints OpenAI's API has exposed since GPT-5.6, placing them after stable request layers and moving variable setup such as skills, subagents and environment info into a "phantom user message." That cut cold cache misses by 20%. Separately, numbering only every tenth line in file reads reduced cache-read tokens 1.6%.
Cursor said it relied on A/B tests over real traffic rather than evals, which it argues skew toward hard problems. The published material on subagent delegation is incomplete in the signal reviewed.