Could AI agents truly make sophisticated software development accessible to almost anyone? Codeium is betting on it, having attracted hundreds of thousands of users to its agent-native Windsurf IDE. There are opportunities galore in science, from education to automating research workflows. In terms of the former, imagine instructing an AI coding partner to build a complex scientific visualization – perhaps an interactive solar system model. Codeium has an animated demo of this on its website, which shows a visualization of planets orbiting around the sun. That same interactive power could allow a biologist to visualize protein folding simulations or help a chemist create custom dashboards for analyzing spectroscopy data in real-time.
By embedding AI deeply into the coding process, the startup aims to dramatically lower the effort required for custom software. “We have taken the activation energy of building custom logic and custom scripts and custom apps down to near zero,” claims Anshul Ramachandran, Codeium’s Head of Product and Strategy, pointing to a future where scientists and business users can build complex tools with ease.
Anshul Ramachandran
For Ramachandran, who holds degrees in both chemistry and computer science from Caltech, the potential impact extends beyond traditional software engineering. He points to internal experiments where Codeium gave its agentic software tool to “every single non-technical person at the company.” He continued: “I’m not talking about people who know Python—they just don’t know how to code at all. And we said, ‘Here’s a tool. Can you build whatever you need to build for your job?'”
Anshul notes that the results in the months that ensued were “very interesting.” Codeium’s partner leader, for instance, built his own partner portal using Windsurf, while the revenue operations lead created a custom quoting tool, and a marketing team member developed an internal survey tool. Ramachandran added that “next thing you know, three months later, we’ve canceled about half a million dollars of internal B2B SaaS spend at the company.”
When internally we were doing research, we realized what we’re doing is combining the human-loop aspects of the co-pilot systems like ChatGPT with the capacity to be independently powerful like agents. That’s magical and intuitive.
Empowering domain experts
While that showcases one end of the user spectrum, Ramachandran places “the sciences and people who are Python-literate, [who] do a lot of data science… somewhere in between” professional developers and complete novices. For this group, already familiar with tools like Pandas and NumPy, Codeium taps the considerable publicly available information these models were trained on, enabling users to “iterate incredibly rapidly.” He notes observing friends in PhD programs using Windsurf effectively: “Instead of working in a Jupyter notebook, they’ll just go to Windsurf,” asking the agent, Cascade, to generate specific reports or analyses from their data sources.
The key advantage, Ramachandran emphasizes, lies in bypassing tedious manual steps. “Forget about trying to do a lot of cell manipulation in Jupyter—it’s like, ‘Here is exactly the script that will do exactly what you need to do.'” Crucially, the agent doesn’t just provide code; it assists with execution. It will “tell [you] how to run the script, how to download the dependencies. It’ll walk through all of those things for you because it’s an agent,” providing specific commands rather than just instructions. The result, he argues, is that researchers can generate highly specific outputs much faster than before, enabling them to “build exactly what you need in that moment of time, from scratch, instantly. Hundreds, thousands of lines of code in a matter of minutes.”
From code generation to a collaborative partner
While the ability to generate code rapidly can be powerful, it isn’t necessarily a time savings if the thousands of lines of code an AI generates take up the better part of two weeks to debug and overhaul. Writing successful code isn’t just about getting it to run; it needs to be correct, maintainable, and fit the user’s actual requirements. Ramachandran explained that Codeium deliberately moved away from early agent concepts where users defined a task and waited passively for results. “Pre-Windsurf, the idea of an agent was: you have a task… send it off… and 20-30 minutes later, it’ll give you a result,” he explained. “We thought that approach was slightly incorrect… are you really even saving time?”
Instead, Windsurf was built around what Codeium calls “collaborative agent flows.” The AI works within the IDE (based on the open-source VS Code) in real-time, explaining its steps as it goes. Ramachandran emphasized the importance of “seeing what it’s doing in real-time.” He added that users can “correct it” at any point, and crucially, “the agent will be aware of the edits that you’re making.” To cite the solar system example mentioned earlier that shows up on Codeium’s website, the user can not only see the visualization of planets orbiting around the sun materialize in a preview window, but also observe the agent’s step-by-step reasoning and code generation in an adjacent panel. If users want to modify the planet’s appearance or adjust the orbital speed, they can provide feedback or edit the code directly. The agent, in turn, can incorporate these changes or even proactively identify and fix errors it introduces, as demonstrated when it detects a syntax error in the demo and proceeds to correct its own code.
While it is often possible to build apps from scratch with little or now software development experience, Ramachandran underscores that it is still helpful to understand the languages the agent is working in. This enables a tighter feedback loop. “It’s very similar to how you or I would peer program with someone else,” he said. “We wouldn’t try to scope out a bunch of work and pass it on. We’d do some work, pay attention, then the other person would do some work, passing back and forth.”
If you’re using an agent, and the agent does something wrong, I’ll be the first to say that having that literacy in code helps you review it and see where things are going wrong. Reviewing code is often a lot easier than writing code.
This interactive model, where “the human and the AI [are] working on the same timeline of work,” not only mirrors human peer programming, but also allows for iterative refinement with accelerated timelines. That is, it aims to prevent the AI from prioritizing quantity of code over quality based on users’ requirements. “We build a platform, but I’m most excited about seeing how people use it. Every day I hear new stories of the stuff that people have been able to build. And I don’t think we’ve even scratched the surface yet.”
[Image courtesy of Codeium]
