Scott Wu on Cognition's Windsurf Deal and the Future of AI Coding

Harry Stebbings welcomes Cognition co-founder and CEO Scott Wu to discuss the company's acquisition of Windsurf and what AI agents mean for software development.

Wu says he learned of Google's licensing arrangement and the departure of Windsurf researchers on Friday, at roughly the same time as the public. He contacted Windsurf CEO Jeff Wang that evening.

According to Wu, the teams reached a verbal agreement on Saturday, worked through legal details on Sunday and signed on Monday. He describes the speed as unusual even for AI.

Wu says the appeal was the combination of Windsurf's product, customers, code, data and intellectual property, along with its go-to-market team. He saw a fit with Cognition's agent work.

The process required balancing many constituencies. Wu describes the remaining team's loyalty to its customers and colleagues as an important reason a deal could come together.

Wu says Windsurf had been considering other options, including raising money independently or entering a partnership. His account of the founders' notice period is presented as his understanding, not established fact.

Stebbings asks whether Google left valuable assets behind. Wu agrees that the product and team remaining at Windsurf had considerable value for Cognition; that is his assessment of the transaction.

The conversation turns to the premium paid for AI talent. Wu argues that exceptional researchers can have outsized impact, although the number of people he believes can drive frontier progress is an estimate.

Asked whether applications can compete with model providers, Wu says the model is only part of the product. Context, tooling, domain knowledge and integration into a user's workflow can all matter.

Cognition works with model providers, including Anthropic. Wu says the relationship between model companies and applications involves both partnership and potential competition.

A product can use a general model while building its own systems around it. Wu sees opportunities for application companies to improve the reliability and usefulness of an agent in a particular environment.

Stebbings asks how much revenue is paid to model providers. Wu declines to provide a percentage, and the discussion does not establish Cognition's margins.

Wu argues that reinforcement learning and domain-specific feedback can improve performance on specialized work. His optimism about solving benchmarks is a view about future capability, not evidence that every task is already solved.

He also believes current models are useful enough that developers can already make substantial changes to their workflows, even without assuming further model advances.

Wu cautions that the percentage of code written by AI is a crude measure: different lines have different importance, and it is difficult to define where AI assistance begins and ends.

He estimates AI tools can make developers roughly 1.5 to 2 times faster today. He predicts much larger gains, potentially 10 times, over the next few years; neither figure is presented as a controlled study.

Cheaper software creation could increase demand as well as productivity. Wu expects people to build more tools and improve products that would otherwise be too expensive to maintain.

Reliable agents need more than code generation. Wu points to understanding the repository and environment, running tests, observing results and using feedback to correct mistakes.

In his view, human work will increasingly include specifying goals, making architectural choices and reviewing outcomes, while agents take on more implementation. He presents this as a direction of travel, not a guarantee.

The value of an agent is tied to time saved, but that value is hard to price precisely. Wu describes Cognition's ambition to charge far less than the value of the work completed.

To handle harder tasks, an agent needs context from prior work and the ability to navigate a large codebase. Wu stresses the importance of knowing how a product actually behaves.

A useful system also needs to gather information, use tools and learn from feedback. These surrounding capabilities can make the difference between a compelling demo and a dependable workflow.

Wu describes the context an agent needs to understand a codebase, test a front end and decide whether its changes work as intended.

These practical details are easy to underestimate. Wu says building a dependable coding agent requires attention to how software engineering works in real teams.

Stebbings questions Cognition's public profile. Wu describes how engineering teams bring Devin into their workflows, including through Slack and Linear.

Wu says people are still becoming familiar with agents and predicts that awareness will increase substantially over the next six to twelve months.

Discussing growth, Wu says usage across self-service and enterprise increased five to tenfold from January to July. The figures are his account, not independently verified.

The pair compare Cognition's trajectory with Lovable, and Wu says it is roughly comparable without confirming an exact ARR. Stebbings discloses an investment in Lovable; Wu leads Cognition and has a commercial interest in its growth.

Stebbings argues that telling a clearer story about growth could help the company recruit and communicate its progress. Wu discusses the trade-offs around publicity.

They turn to the future of IDEs and autonomous agents. Wu thinks it is too early to know which interface or company will dominate.

Wu calls the next generation of human-computer interaction an important problem: users may increasingly express intent instead of directing every individual edit.

He predicts rapid progress in interfaces and capability, but acknowledges that how the categories develop remains uncertain.

Combining Windsurf and Cognition brings together teams working on adjacent parts of the software development process, in Wu's account.

Wu describes an approach that combines synchronous planning, asynchronous agent execution and human review. He sees a role for both the editor and the agent.

The integration raises product questions about how people will move between direct editing and delegating work. Wu says there is more to learn as the teams build together.

In the quick-fire round, the discussion turns to AI valuations and whether the market resembles a bubble.

Wu believes aggregate value for foundation-model companies and applications could rise over five years, while individual companies may disappoint. This is his investment outlook, not a guarantee.

Wu argues that reinforcement learning can improve agents when they receive useful feedback about whether their work succeeded.

He contrasts feedback with simply imitating examples. Different professions need different evaluations, and defining a reliable test can be difficult.

Wu is bullish on applying these methods across domains, but his broad claims about their potential are forecasts rather than demonstrated results in every profession.

Asked to choose between investing in Anthropic and OpenAI at the valuations cited by the host, Wu sees potential in both and ultimately mentions Anthropic. Cognition works with Anthropic.

Wu predicts a small number of major model providers will remain competitive. Stebbings offers a different view; neither can know the eventual market structure.

Wu says the Windsurf consideration included cash and stock but declines to disclose the proportions.

The pair discuss Cognition's acquisition announcement and the pressure of preparing it while lawyers and teams were finalizing the deal.

Wu says the emphasis in AI training has shifted toward carefully curated data for a specific use case, rather than simply acquiring more data.

He describes Cognition's Kevin 32B model and says reinforcement learning improved its performance on a GPU-kernel benchmark. That is his account; a benchmark alone does not establish broad real-world performance.

Wu recommends studying computer science, arguing that problem solving, first principles and understanding systems will remain valuable as AI changes coding.

Asked which personal trait he values, Wu names his ability to remain composed during stressful situations, including the previous week's negotiations.

Stebbings suggests telling clearer stories about how Cognition's products empower users and how the company is growing.

Wu says more companies are now building agents, making him more willing to discuss Cognition's work publicly. The conversation ends with the pair considering the value of a stronger brand.