On 25 February 2026, Anthropic announced the acquisition of Vercept, a nine-person Seattle startup most people outside the AI world had never heard of. No financial terms were disclosed. No flashy press release with a quote from a celebrity investor. Just a quiet announcement that a small team with a very specific skill set was joining Anthropic, and that their product, a Mac-based computer agent called Vy, would be shut down within 30 days.

That’s exactly the kind of story that gets buried under the bigger headlines. It shouldn’t.

What Vercept Actually Built

The way most AI computer use works today is that the AI reads code. When Claude or any other agent takes control of your machine to complete a task, it navigates the file system, calls APIs, and interacts with software through its underlying structure. That works reasonably well for well-built applications with clean interfaces. It falls apart the moment it encounters anything that doesn’t have a neat code layer underneath, which is most of the software that most people use most of the time.

Vercept took a different approach. Their Vy agent used what they called vision-based UI interaction. It didn’t try to read the code. It looked at the screen. It recognised what was there, the same way you would, and acted on what it saw. If a dialog box appeared unexpectedly, it could handle it. If a button moved, it could find it. If software it had never used before looked intuitive, it could work with it. That is a fundamentally different kind of capability.

I tried to explain this on Prompt Fiction this fortnight using the TeamViewer analogy. When you give someone a remote support session, they take control of your mouse and keyboard and see exactly what you see. The operating system doesn’t know it’s a different person doing it. It just looks like a human using the computer. Vercept had built the equivalent of that, but for AI. And critically, they had done it in a way that was beginning to outperform tools from OpenAI, Google, and Anthropic themselves on UI benchmarks.

Sources: Anthropic blog (25 Feb 2026), GeekWire (25 Feb 2026), TechCrunch (25 Feb 2026)

Why Anthropic Wanted This Team

The numbers tell part of the story. Anthropic confirmed in their announcement that Claude’s computer use scores on the OSWorld benchmark, which is the standard test for AI operating computers, have gone from under 15% in late 2024 to 72.5% today. That’s not an incremental gain. That is a step change over a relatively short period. And they announced the Vercept acquisition in the same breath.

The three co-founders joining Anthropic, Kiana Ehsani, Luca Weihs, and Ross Girshick, came from the Allen Institute for AI. They had spent years working specifically on the question of how AI systems can see and act within the same software that humans use every day. That expertise is precisely what Anthropic needs to close the remaining 27.5% gap and get Claude to genuinely human-level performance on real computing tasks.

What’s also notable is what the acquisition says about the broader competitive landscape. Vercept’s co-founder Matt Deitke had already left the company in mid-2025, reportedly with a $250 million offer from Meta’s Superintelligence Lab. OpenAI hired Peter Steinberger, the creator of OpenClaw, the week before. Every major AI lab is currently racing to secure the people who are building the tools that will let AI operate computers autonomously. Anthropic just picked up one of the best teams in that specific niche.

Source: Dataconomy (26 Feb 2026)

What This Means for the Rest of Us

The practical implication of all of this is something I got a small glimpse of this week with the Claude plugin for PowerPoint. I had Claude building slides in my existing theme while I worked on a different slide simultaneously. Two workers, one document, nobody in anyone’s way. I was producing work at a rate I hadn’t managed before, and when I stopped and thought about what was actually happening, it felt genuinely new.

That’s today, with AI that still has limitations and can still be clunky. Now imagine AI that can see anything on your screen, open any application it needs, handle unexpected pop-ups and errors without stopping, and operate across your entire digital environment the way a capable human colleague would. That’s not science fiction. That’s the direction this acquisition is pointing, and based on how quickly the benchmarks have moved, it’s closer than most people realise.

The question I keep coming back to at Techosaurus, and the one I’d encourage businesses to start thinking about now, is: what are the tasks in your organisation that are essentially coordination? Where a person receives information, processes it slightly, and passes it somewhere else? Those are the tasks that will be handed to agents first. Not because AI is going to replace the people doing them, but because those people are going to use AI to do it in a fraction of the time, freeing them up for the things that actually need a human.

The Bit That’s Easy to Miss

There was a side story in the Vercept coverage that I think is worth noting. Oren Etzioni, one of Vercept’s co-founders and an investor, publicly described the outcome as “throwing in the towel.” He was disappointed. He thought the company had the traction to go further independently. A very public dispute followed between him and the lead investor, Seth Bannon, on LinkedIn.

That kind of thing usually gets treated as founder drama and ignored. But I think it tells you something important about where we are in this industry right now. Even a promising startup with serious technical differentiation, backed by Eric Schmidt and Jeff Dean, outperforming the big labs on their own benchmarks, can find itself folded into one of those labs within a little over a year of its seed round. The resources required to compete at the frontier level are so vast that the realistic exit for most of the best technical teams is acquisition. Which means the interesting question isn’t who’s winning the AI race. It’s which large platforms are successfully bringing in the builders.

Anthropic’s second acquisition in three months suggests they’re getting better at that. Watch this space.

I discussed the Vercept acquisition and what it means for Claude’s computer use capabilities on the latest episode of Prompt Fiction. Listen to Chapter 12, Part 1 here.

Scott Quilter | Co-Founder & Chief AI & Innovation Officer, Techosaurus LTD