Going Beyond Code, Superintelligent AI Agents, And Why Taste Still Matters

Truell says Cursor aims to replace today’s labor-intensive coding with a more productive way to build software while preserving control over how it works and looks. AI already writes 40% to 50% of code in Cursor, but professional developers still review the output. He expects AI suggestions and delegated agent tasks to become much more useful over the next six months to a year.

Truell identifies context, continual learning, and long task horizons as obstacles to more capable coding agents. A codebase with 10 million lines could contain about 100 million tokens, making it hard to provide models with useful, affordable context. Agents also need to run software, inspect results, and use tools such as Datadog logs.

Truell argues that engineers will still need to decide what is worth building and how it should work, including its nonvisual logic. He describes much of programming as translating human intent into detailed instructions for computers. As AI fills in more implementation details, he expects that translation work to shrink while human judgment about the product remains important.

The founders first built AI tools for computer-aided design, but set that work aside because the team was more interested in coding and the available data and models were not ready for 3D. They returned to coding, released Cursor after about three months of development, then spent roughly a year refining it. They chose to build an editor rather than an extension because they expected AI to change how people work with code.

Cursor tracked revenue and paid power users who used AI four or five days a week, rather than focusing only on daily or monthly active users. Truell says the team paired internal use and product experimentation with retention and usage metrics. He also sees product feedback, such as which suggestions people accept, reject, or later correct, as useful for improving the product and its models.