Improved token efficiency for longer agent runs

Cursor reports reducing agent token costs by 7% without lowering quality by optimizing its agent harness, which controls request assembly, context reuse, and subagent delegation. Key changes included cutting roughly 66% of the system prompt as models became more capable, dynamically loading infrequently used built-in and MCP tool definitions rather than including them in every request, and retaining only commonly needed tools in static context, reducing their description tokens by 60%. The company also improved prompt-cache reuse by placing explicit cache boundaries around stable instructions and tools, lowering cold cache misses by 20%, and reduced file-reading overhead by numbering only every tenth line while preserving citation usefulness. For longer tasks, Cursor adjusted subagent guidance and model selection to gain the context-isolation benefits of delegation while limiting duplicated work and unnecessary coordination. It plans to continue measuring context growth and applying similar harness-efficiency methods to Grok Bot.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.