Varun Mohan on Windsurf, Coding Agents and Startup Focus

Harry Stebbings interviews Windsurf co-founder and CEO Varun Mohan about pivoting, AI coding tools, competition and the future of software work. The episode title refers to an acquisition, but the interview transcript does not establish a transaction's terms or completion. Mohan's product claims and predictions come from the leader of a coding-tool company.

Mohan says a useful startup idea can be non-obvious, but most unconventional ideas are simply bad. Founders need optimism to try something and realism to change course when evidence does not support it. Being persistent, he says, does not reward a team for continuing to do the wrong thing.

His company began as Exafunction, working on GPU virtualization. The team correctly expected GPUs to become important, he says, but overestimated the diversity of model architectures and concluded it would be hard to differentiate its infrastructure. It pivoted toward code AI; this is his interpretation of a rapidly changing technical market.

Stebbings asks what idea Mohan has been too attached to. Mohan says, in retrospect, the company could have pivoted earlier from GPU infrastructure and could also have built Windsurf sooner. Customer funding difficulties in 2022 and changing technology helped prompt the first move; he treats the timing as a lesson rather than proof of what an earlier pivot would have achieved.

Asked whether being first matters when competitors can copy quickly, Mohan says it signals a willingness to disrupt one's own product and allows faster learning. Stebbings suggests followers can avoid another firm's mistakes; Mohan replies that outsiders cannot see the internal experiments that failed.

As an example, he says the team tested a code-review product through a browser extension and a separate site before shipping a later version. An early coding agent was also inadequate; work on understanding codebases and improvements in models eventually made the Windsurf product viable in his account. His assessment of its value is a company claim.

Stebbings asks whether more funding helps founders run more experiments. Mohan says a Green Oaks round for the original GPU business gave the company cash to pivot and launch its earlier Codeium extension free. He says capital only helps if the organization is willing to change direction quickly, and recalls the team moving to the new idea over a weekend despite revenue from the old one.

Mohan argues a startup usually succeeds by doing one thing exceptionally well. When a new direction is materially stronger, he says, teams should concentrate resources rather than split attention between unrelated products. That can mean difficult conversations with investors and customers about ending support for the old business.

He says non-developers already use Windsurf and cites an employee who built internal partnership and quoting tools that replaced more than $500,000 of sales-tool spending. The company example is his account; the interview gives no independent accounting or full cost comparison. Windsurf's main focus, he says, remains professional developers and large codebases.

Stebbings asks whether developer tools and products such as Lovable and Bolt will converge. Mohan predicts that tools aimed at non-developers will gain more controls, while better codebase understanding will make app creation easier for nontechnical people. He sees a present trade-off between serving that use case and improving the experience for professional developers.

With unlimited resources, Mohan says, he would run more experiments. He estimates at least half of internal ideas fail and argues that a successful one can justify many failures. Those figures describe his view of internal experimentation, not audited project or financial returns.

He favors small teams of roughly three or four people for unproven ideas, to avoid coordination and disagreement before evidence exists. His rule of thumb is that even an imperfect first version of a great idea should reveal something compelling; only after that should more people be assigned. He cites the early agent prototype as an example.

Mohan says the company looks at progress over time and can pause a project through a top-down decision rather than setting a fixed budget for every bet. In hindsight he wishes the team had improved autocomplete sooner; he says the VS Code interface constrained its early product experience before Windsurf's editor.

Stebbings, who discloses being an ambassador, asks whether AI has made startup moats obsolete. Mohan says durable advantages are hard for young companies to establish and places the greatest weight on speed and accumulated learning. His characterization of Windsurf's talent and defensibility is a competitive claim from its CEO, not an independent assessment.

He argues that Nvidia's continuing hardware and interconnect improvements matter more than treating CUDA alone as an unassailable moat. Stebbings counters that a large company's distribution, brand and resources can create powerful advantages for an acquired product. Mohan maintains that fast learning and execution can still allow startups to win; these are competing strategic interpretations.

Mohan attributes some large-company slowness to weaker urgency than a startup faces. He says Windsurf ships major releases every one to two weeks, a company statement not independently verified here. The comparison does not establish that all large firms are slower or that release frequency alone measures quality.

Windsurf is an in-person company, Mohan says. He thinks colocated engineers can coordinate and pivot faster, though he acknowledges remote companies can work with stronger principles and process. This is a management preference from his experience, not a controlled comparison of remote and office teams.

On brand, Mohan says recognition helps new products reach users but does not excuse slower innovation. He recalls earlier enthusiasm around GitHub Copilot and Devin and argues that coding tools must repeatedly prove their value. Stebbings asks whether rapid shifts should lead to lower valuations; the interview does not resolve that pricing question.

Mohan says enterprise deployments can create switching costs, but acknowledges those costs are not yet like those of established enterprise software. He says continuous improvement matters because customers can still move to a better product. The discussion is a CEO's explanation of retention, not a quantified churn analysis.

He describes Windsurf's own model as comparable to some frontier models on an agentic workload while faster and cheaper for the company, and says it processes hundreds of billions of code tokens per day on its GPUs. The benchmark, price comparison and traffic claims are not independently verified in the interview.

Asked how much use or revenue is enterprise, Mohan says more than 50% of the total is enterprise, without specifying the denominator. He highlights support for JetBrains IDEs as well as the Windsurf editor and cites JPMorgan Chase as a customer. The account does not establish a separately measured enterprise revenue share.

On engineering employment, Mohan says the answer will vary by customer and that the definition of engineer may broaden. He expects some people to build using natural language, while production-critical systems still need people who can inspect and validate details. He predicts fewer people may need to work at the deepest level, but supplies no employment data or timetable for particular firms.

Asked whether product managers will remain, he expects them to take greater responsibility for building and testing their own ideas, supported by AI tools. He says rapid prototypes may precede a detailed design process, while core product experiences still need design work. These are forecasts, not evidence that every product role will change in the same way.

Mohan cautions that website and app prototyping is only part of software engineering: complex databases, infrastructure and large codebases remain. Windsurf aims to help developers across that broader work. His description is also a product-positioning argument by an interested CEO.

He expects asynchronous remote agents to handle some tasks, such as producing a pull request after a user delegates work. But he says quality, latency and how easily a result can be corrected must be considered together. Waiting a long time for a flawed change can make review and repair costly.

Mohan notes that even a result that appears 90% correct may be difficult to fix because the reviewer must identify and understand the rest; he says others argue for 99% correctness. These are illustrative thresholds, not validated reliability standards. He expects simpler tasks to be better suited to remote asynchronous work in the near term.

He says small latency changes affect acceptance of the autocomplete product and discusses possible interfaces for remote agents. A phone may be useful for a one-off request, but reviewing thousands of lines of code on one is hard. He does not claim to know the eventual form factor, and his product metric is a company observation.

Stebbings asks whether existing applications will collapse into databases operated by agents. Mohan doubts Salesforce or Workday will be replaced so simply in the short term because organizations have built complex workflows around them. He also says current agents are not trusted to modify databases broadly without human supervision.

Mohan thinks some investors overestimate agents' present ability, recalling excitement about replacing junior developers. At the same time he predicts capabilities and trust will improve rapidly, potentially within months. Both the near-term assessment and the timeline are his views; the interview does not prove a junior role can or cannot be fully automated.

Over the following year, he expects tools to help more with design, debugging, testing and deployment as agents gain access to logs, databases and other context. He predicts large productivity gains across software work. The tenfold figure is a forecast, not a measured result, and expanded data access also requires appropriate controls.

Asked what Cursor did well, Mohan praises its user interface and experience. He says Windsurf initially looked for a deeper technical breakthrough before building its own editor. He credits both frontier model improvements and Windsurf's own models and interface work for making coding agents usable, without independently comparing products.

Stebbings asks whether model providers resemble cloud providers. Mohan says switching between model APIs is currently easier because they generally retain little state, though he imagines future systems could retain more user and codebase context. He warns providers not to prioritize lock-in over improving models in a fast-changing market.

On commoditization, Mohan says he does not expect one provider to hold an unchallenged lead in a valuable category for years in the near term. He believes scale and capital matter but says new techniques can change the race. This is a forecast made by the CEO of an application that uses and builds models, not proof that all models are interchangeable.

He says model providers can differentiate through specialized APIs—for example, reducing latency for a coding workload—and may also build their own applications. He sees both approaches in the market. The future division of value between model and application companies remains uncertain.

Stebbings cites a prediction that one-person billion-dollar companies will appear soon. Mohan disagrees: he argues competitors with similar tools and more people could put pressure on prices, brand and customer acquisition. His argument concerns competitive dynamics, not a proof that no such company can exist.

Mohan says the most anxious moments came before pivots, when he believed the company could do better but inertia prevented action. After the move from Exafunction to Codeium, he felt freer even though he thought failure was possible, because the team was pursuing something it believed in.

On hiring, he says the company closed large enterprise deals before hiring a sales leader, which helped it understand the role and attract a candidate. He prefers making a hire when the need is real rather than far in advance; he says an unused executive may create work that does not matter. This is a trade-off, not a universal staffing rule.

Asked what he has changed his mind about in management, Mohan says more people do not automatically produce more output. He finds a team with a single clear priority easier to lead than one with the same revenue and five priorities. The discipline to say no, he argues, is difficult but central.

In quick fire, Stebbings discloses that he is an ambassador for investor Neil Mehta; Mohan praises Mehta's analytical flexibility and generosity with time. He says he would value a board member like Scott Cook for experience building a lasting software company. These are personal assessments and the investor relationship is relevant.

Asked about open-sourcing Windsurf, Mohan says opening the front end alone would not reveal most of the logic he considers differentiated in the back end. He thinks it would have limited effect on adoption; that is his hypothetical product assessment. For an 18-year-old, he emphasizes problem-solving and agency over a particular programming language.

Mohan says he could spend less attention on details that do not affect the company's long-term success. He closes with the team's ambition to reduce the time needed to build technology by 99%. That number is a mission, not a demonstrated or guaranteed improvement.