Windsurf Launch

Think back to a moment when you truly felt magic when using technology.

The first time you asked ChatGPT a question? The first time you used a smartphone touch screen? The first time you Googled something specific and got the perfect resource?

Magic happens when the interaction with the technology is simple, seamless, and intuitive, but it is obvious that something very powerful is happening under the surface.

While we have been launching AI products with real value over the last couple of years under the name of Codeium, we have in parallel been quietly iterating on this concept - what does magic look like for a software engineer using AI? How can a developer be seamlessly in the flow while still recognizing the technology’s power?

Our answer is the Windsurf Editor, generally available today.

The Spark

Any magic has some spark of intuition.

We started with the existing paradigms of AI use. Copilots are great because of their collaborativeness with the developer - the human is always in the loop. That being said, to keep the human in the loop, copilots are generally confined to short scoped tasks. On the other hand, agents are great because the AI can independently iterate to complete much larger tasks. The tradeoff is that you lose the collaborative aspect, which is why we haven’t seen an agentic IDE (yet). An IDE would be overkill. Both copilots and agents are powerful, but have generally been seen as complementary because their strengths and weaknesses are indeed complementary.

Our spark came from one simple question - what if the AI had the best of both worlds? What if the AI was capable of being both collaborative and independent? Well, that is one aspect of what makes humans special. Working with that kind of AI could feel like magic.

With a lot of research, we built the foundations of this kind of system, which we are calling AI flows. AI flows allow developers and AI to truly mind-meld, combining the best of copilots and agents.

To understand an AI flow, let’s start with what it has access to: context of knowledge, access to tools, and real time understanding of human actions. AI copilots have gotten better over time with improved context of existing knowledge. AI agents are a step forwards by additionally incorporating access to tools. AI flows allow for true human-AI unity by adding on real time understanding of human actions. The combination of knowledge, tools, and human actions allows the human and the AI to always collaborate on the exact same state of the world, all while still arming the AI to be as independently powerful as possible.

In the world of software development, one might like AI to take over through stretches of boilerplate, but when working on a particularly critical part, would like to not have the AI involved. But after returning to AI assistance, the developer would want the AI to automatically be aware of what they did without the AI. That mind-meld is possible only with an AI flow - an omniscient, collaborative AI agent.

To appreciate the value of incorporating human actions, we need to think about actions in terms of timelines. Before LLMs, we could represent a developer’s work with a sequence of actions on a single timeline - steps of edits, navigation, research, etc.

With LLMs, we started invoking ephemeral AI copilots to answer questions or perform scoped tasks, such as to answer a chat message or perform an inline code refactor, and with access to existing knowledge bases, these copilots could give relevant results. These AI copilots have short timelines, ending once the task is complete.

The value of AI agents is to expand the capacity of these copilots to tackle larger problems by arming the AI with tools and iterative abilities. These AI agents can execute a divergent timeline of work, with results that come back to the developer to review and merge back into their simultaneously progressing timeline of work.

But why should AI be scoped to a problem or caught up to speed? Why add work on the review end because the developer and AI timelines have diverged? What if the AI just understood the developer and provided deep assistance whenever valuable? Those are the motivations for having the AI and developer timelines flow together, and that is only possible if the AI is fast enough to real time incorporate the developer’s trajectory or actions.

This was the spark. Not AI copilots, not AI agents, but AI flows.

Now, we needed to build the magical experience, a new surface that could natively expose this spark.

The Magic

With all of the talk of flows, magic, and power, the name “Windsurf” spoke to us. Windsurfing perfectly captures the combination of human, machine, and nature in an activity that looks effortless, but takes an intense amount of power. We decided to utilize “Windsurf” as the identification of this new surface, a brand new editor, the Windsurf Editor.

And then the aquatic analogies kept coming, starting with the core flow - Cascade.

Cascade is the flow evolution of Chat, and you will find Cascade in the side panel instead of Chat. On face value, it may look like a familiar conversational surface. That’s the “intuitive” part of the magic. Once you start using it, you will notice the “power” part - the synthesis of deep reasoning of your existing codebases (the knowledge), access to a vast array of tools, and omniscience of the actions that you are taking independent of invoking the AI.

Just watch.

What makes Cascade insanely powerful is not just the breadth across knowledge, tools, and human actions, but the depth within each axis:

• Knowledge: This is where our multi-year work on building state-of-the-art context awareness systems that can parse and semantically understand complex codebases comes into play. If we weren’t really good at this, we wouldn’t be fortunate enough to be able to work with some of the largest and most technically complex companies such as JPMorganChase, Dell, Anduril, and Zillow.

• Tools: Cascade’s tools include making edits, adding files, grep, listing files in a directory, and even code execution. On top of this, Cascade comes with proprietary tools such as Riptide, which is the technology that underpinned our research breakthrough that was covered by the press a few months ago. It is an LLM-based search tool that can rip through millions of lines of code in seconds with 3x better accuracy than state-of-the-art embedding-based systems, all with highly optimized use of a large amount of compute.

• Human Actions: There are a lot of different granularities at which you can capture this information, but it is very easy to either have too little or too much information. Either you miss actions core to determining user intent or you have too much noise. We won’t give away the magic sauce here, but we have done a lot of work on checkpointing, information compression, and more in order to make Cascade feel like an infinite stream of joint consciousness between human and AI.

So, if Cascade is the flow evolution of Chat, what about Autocomplete?

Well, you may have already used the flow evolution of Autocomplete - it’s Supercomplete! Predicting the next intent, not just the next text at the cursor position, is only possible with flows and the ability to reason over human actions.

While flows were the motivation behind the Windsurf Editor, Cascade is not the only AI capability you will have access to on day one. The Windsurf Editor was built around our overall AI engine to assist you however you write code, whether it’s Cascade or in the text editor. As you use the text editor, you will get all of the features you have come to love from the Codeium extensions.

You can enable the fast Autocomplete mode and get blazingly fast and contextually aware Autocomplete suggestions to finish your thoughts. Also included is Supercomplete, the modality that predicts the edits corresponding to your next intent, prompting you with multi-cursor edits to keep you in the flow. Command has a brand new UX that makes it easier and faster to perform large modifications, full file generations, and follow ups. We have even brought Command into the terminal so that you never have to Google that random terminal command again. These are just a handful of the AI features infused into the Windsurf Editor.

The Windsurf Editor is built to keep you in flow state. Instant, invaluable AI developer assistance where you want it, when you want it.

The Tactical

With all of the talk about magic, we have enough self-awareness to know that one of the reasons we have been successful so far is that we are brutally honest with ourselves on the realities of both the technology and the user.

We did not go into building an editor until we realized the magic of flows and Cascade.

That being said, we also were honest with ourselves that we did not have to build the editor entirely from scratch to expose this magic, so we forked Visual Studio Code. We are fully aware of the memes about people forking VS Code to create “AI IDEs,” but again, we would not have built the Windsurf Editor if extensions could maximize the potential of our vision. We have been an extension-first company, and still recognize that people really like the editors that they have, especially within our enterprise customer base. So, our Codeium extensions are not going anywhere, and we are going to continue to improve them to the max of their capabilities. Even some flow capabilities like Supercomplete are doable within extensions, and so we will build them in! The only difference with the Windsurf Editor is that we now have a surface where we are truly unconstrained to expose the magic as it evolves.

The next tactical piece is cost. Our infrastructure expertise has been the secret sauce behind a number of the loved aspects of our Codeium extensions, from the crazy low latencies to the generous free tier (it’s not a financially irresponsible option for us due to our industry leading serving costs). But even for us, serving this magic at its full potential is a meaningful jump up in operating cost. So while the Windsurf Editor itself and a lot of the Cascade capabilities will be free, the full magic will only be available on paid plans in the long run. That being said, for the next week, we are going to be giving two weeks of the full experience for free to any individual using the Windsurf Editor.

Finally, for our enterprise customers, we cannot wait to give you access to the Windsurf Editor, but just like with everything we build, we want to make sure everything is stable at scale before rolling it out to you. Again, that is the practicality that comes from running an AI company with thousands of enterprise customers in production.

The Future

We took a poll internally on the engineering team - are you more excited about this launch or what we will be shipping shortly in the “first wave” of updates to the Windsurf Editor? The latter was a near unanimous decision. So let’s give a glimpse to why.

We plan on giving Cascade access to more knowledge - expand the reasoning across not just the codebases, but existing tickets, docs, Slack messages, you name it. If there is a store of knowledge that is useful to a developer, it will be useful to Cascade. This will allow Cascade to do even more impressive tasks on even larger and more complex knowledge bases.

We also plan on giving Cascade even more powerful tools. What if we used developer behavior to understand semantically what files are related to each other, even if syntactically there are no code paths or signals that they would be connected? For example, we can understand that certain frontend code is changed whenever certain backend code is modified, and perhaps even a yaml file for deployment. This clustering would be a powerful tool to retrieve relevant information.

We will start to take advantage of this timeline representation of flows, by predicting various branches of what the developer wants to do in the future given the behavior we have observed and iterated with thus far. By predicting more of your intent, we will create a flow version of the Command functionality. These intents can appear as an optional selection for a new Cascade, but we have other ideas for novel user experiences with such proactive capabilities.

We are starting to piece together what it would look like if the Windsurf Editor itself was a collaborative surface between humans, and if Cascade is the place where AI can participate in the day-to-day collaboration between peers. For the first time, we have a smart enough “machine” that would add value to and benefit from being part of the conversation.

And then of course, how can we continue to extend the flows across more key work surfaces of software development? If the AI is in lock step with you through documentation, work planning, code review, and more steps that aren’t just purely coding, then the value of the AI will simply compound.

This launch is just the beginning.

Closing Thoughts

We don’t mind if you call the Windsurf editor the first agentic IDE, the first native surface for developers to collaborate with AI, or simply how we like to think about it - tomorrow’s editor, today.

We just hope you take the opportunity to experience magic with technology one more time.