Varun says the company began as Exafunction, helping customers run GPU workloads, and had reached a couple million dollars in revenue. In 2022, the team concluded that transformer models could make its infrastructure business easier to commoditize, so it pivoted to coding tools over a weekend. Eight people began building Codeium, which shipped within two months.
Betting On AI Agents, Pivoting In 48 Hours, And The Future of Coding
Speaker
Varun Mohan, co-founder and CEO of Windsurf, joins Y Combinator’s The Light Cone to discuss the company’s pivots, AI coding agents, and how software development may change.
After growing its enterprise business, the team shifted toward agents as models improved, then built the Windsurf IDE because it felt VS Code limited the experience they wanted. Fewer than 25 engineers shipped it in under three months. Varun also describes supporting multiple IDEs and large company codebases as important for serving customers such as Dell and JPMorgan Chase.
Varun says Windsurf bet early on agents that could understand a codebase and make changes, rather than relying mainly on chat and autocomplete. For code retrieval, the team combines keyword search, vector search, syntax parsing, and GPU ranking to find relevant context. Its evaluations use code changes and tests to measure retrieval, intent understanding, and whether the resulting code passes.
Varun says developers at Windsurf increasingly give agents tasks before editing code themselves, including repetitive work and deployment workflows. He recommends committing frequently so agent changes can be reverted safely. He expects AI to reduce boilerplate and give engineers more time to test product ideas, while broadening software creation to include nontechnical builders.
Varun describes Windsurf’s opportunity as improving on what foundation models can do, while recognizing that the baseline keeps moving. He points to code migrations and automated bug and alert resolution as areas where specialized products could deliver value. His advice to founders is to change their minds quickly when evidence undermines their ideas.
Takeaways
- Treat a startup thesis as something to test against changing evidence, and pivot quickly when it stops holding up.
- Choose a pivot that the team is motivated to pursue, then commit to it together.
- Use evaluations tied to real outcomes, such as whether code changes pass tests, to guide product complexity.
- For agent-assisted coding, commit frequently so you can recover from broad or incorrect changes.
- Look for repetitive engineering work where a specialized tool can deliver measurable value.
Mentioned in this talk
Search inside this talk
What was claimed 25
- foundedVarun Mohan founded
WindsurfCo-founder and CEO“the co-founder and CEO of Windsurf”
YouTube Verified3 sources· Talk, Web· archived
- ceo ofVarun Mohan CEO ofWindsurfCEO
“the co-founder and CEO of Windsurf”
YouTube Unverified1 source· Talk
- metricCodeium: Well over a million developers had used the product and it had hundreds of thousands of daily active users.
“well over a million developers use it, and it has hundreds of thousands of daily active users”
YouTube Unverified1 source· Talk
- previously atVarun Mohan previously at
ExafunctionCo-founder“We didn't start as Windsurf; we started as a company called Exafunction.”
YouTube Unverified1 source· Talk
- uses
“we were early adopters of GitHub Copilot”
YouTube Unverified1 source· Talk
- built
“everyone started working on Codeium, the extension product, immediately after”
YouTube Unverified1 source· Talk
- feature
Microsoft Copilot: AI coding assistant used by the founders as an early signal for coding tools.
“we were early adopters of a product called GitHub Copilot”
YouTube Unverified1 source· Talk
- metricCodeium: The earliest version was materially worse than GitHub Copilot; its main advantage was being free.
“The only difference was that it was free.”
YouTube Unverified1 source· Talk
- launchedCodeium: VS Code extension
“within two months, had shipped the product”
YouTube Unverified1 source· Talk
- metricCodeium: By the beginning of 2023, autocomplete capabilities were much better than Copilot's.
“our autocomplete capabilities were much better than Copilot's”
YouTube Unverified1 source· Talk
- featureCodeium: Autocomplete models trained for filling code in the middle of a line
“our model could fill in the middle of code”
YouTube Unverified1 source· Talk
- metric
JPMorgan Chase: Tens of thousands of developers used the product internally at customer companies.“These companies now have tens of thousands of developers using the product internally.”
YouTube Unverified1 source· Talk
- usesJPMorgan Chase uses Codeium
“companies like Dell and JPMorgan Chase became customers”
YouTube Unverified1 source· Talk
- feature
Claude 3.5: Its arrival enabled the company to make its agent capabilities work effectively.“in the middle of last year, that completely changed with the advent of Claude 3.5”
YouTube Unverified1 source· Talk
- launchedWindsurf Editor: AI coding IDE
“in the middle of last year we decided we needed to build our own IDE”
YouTube Unverified1 source· Talk
- metricWindsurf Editor: Shipped less than three months after the project began, across all operating systems.
“we shipped Windsurf less than three months after starting the project”
YouTube Unverified1 source· Talk
- metricWindsurf: The engineering team had fewer than 25 people when Windsurf was shipped.
“The engineering team was probably still less than 25 people.”
YouTube Unverified1 source· Talk
- featureWindsurf Editor: Agent-centered editing with codebase understanding and rapid code changes
“We took a very opinionated stance that agents were where the technology was going.”
YouTube Unverified1 source· Talk
- usesCodeium uses vector database
“a vector database is one tool in the toolkit”
YouTube Unverified1 source· Talk
- uses
“We use a combination of keyword search, RAG, abstract syntax tree parsing”
YouTube Unverified1 source· Talk
- usesCodeium uses Vector RAG
“We use a combination of keyword search, RAG, abstract syntax tree parsing”
YouTube Unverified1 source· Talk
- featureAgent evaluation: Measures retrieval accuracy, intent accuracy, and whether generated changes pass tests.
“You can measure retrieval accuracy, intent accuracy, and test-passing accuracy.”
YouTube Unverified1 source· Talk
- featureGit: Worktrees may let multiple agents operate on separate versions of a repository.
“Perhaps you can have many agents working on different worktrees.”
YouTube Unverified1 source· Talk
- metricWindsurf: A non-trivial chunk of users had never opened the editor and used Cascade instead.
“a non-trivial chunk of our developers have never opened the editor”
YouTube Unverified1 source· Talk
- integrates withWindsurf Editor integrates with Cascade
“They just live in Cascade, our agent.”
YouTube Unverified1 source· Talk
