Summary

Cursor’s differentiation rests on shipping useful automation before ambitious demos. Asif says the team withheld roughly three agent prototypes because they were not useful enough for daily use, while pushing early inventions such as next-edit prediction and working toward repository-wide edits. The governing rule was to be “the most useful product at any moment in time,” not to overpromise at the frontier.
Usage scale feeds a product-and-model flywheel. Cursor’s custom Tab model handles about 100 million requests a day, generating data for workflow improvements and model training. Asif describes Apply as a loop in which a larger model supplies data for progressively smaller, faster models, until edits to a 1,000- or 2,000-line file can feel “effectively instant.”
The winning coding interface may combine high-level automation with permanent fine-grained control. Asif expects Cursor a year from now to look materially different from VS Code, perhaps letting humans edit a pseudocode-like representation while models work underneath. Yet developers will still want to “go in and toggle some of the indices” when direct editing is faster or safer.
Context—not merely benchmark intelligence—is a central constraint on agent quality. Early 4K-8K windows could barely contain a current file; reliable 50,000-60,000-token requests now let models search repositories and sustain broader edits. The remaining prize is architectural understanding: identifying “what is really going on” may require reasoning across billions or tens of billions of tokens.
Cursor’s responsiveness sits atop substantial infrastructure and fixed GPU capacity. Its indexing infrastructure processes billions of files per day, while some larger-model inference can consume tens of thousands of tokens “per keystroke per person.” Lukas highlights the lack of a great serverless-GPU option, making capacity allocation a continuing infrastructure problem.
DeepSeek was an operating advantage for Cursor before it became an investor shock. Cursor had already run DeepSeek V2 for “eight or 10 months—probably 12,” scaling it to hundreds of millions of calls after judging its pretrained models capable, economical, and unusually trustworthy on evaluations. Asif also expects DeepSeek V3 to be a strong base for custom applications.
More autonomous coding is coming, but Asif expects adoption to feel gradual rather than discontinuous. Models must progress from coherence over tens of tool calls to hundreds or thousands; until hands-off systems become genuinely useful, Cursor will “just wait until it gets good enough.” Over one to two years, faster prototyping could move organizations toward “less talking, more coding,” replacing debates over hypothetical designs with arguments over working artifacts.

Deep dive

Not yet available upstream; scheduled sync will retry.

Mark Zuckerberg — AI will write most Meta code in 18 months 112: Talking Embodied Intelligence with 千寻’s 高阳: How Someone Who Looks Like a Robot Builds Robots That Look Human