All the major AI labs are developing AI software agents that can operate a computer just like a person, visually parsing pixels, moving the mouse, and pressing keys. These are Computing User Agents (CUAs), and I wrote in-depth about them last week.
Today’s best CUAs are able to complete about 45% of the tasks in the popular OSWorld benchmark, up from just 6% when the benchmark was created sixteen months ago. What happens when they reach 100%?
In this post, I’ll explore what these benchmarks really measure, what they leave out, and how to prepare for the moment that AI UI execution becomes a solved problem.
The OSWorld benchmark defines 369 real desktop tasks: file management, web browsing, multi-app workflows, and so on. Human testers are able to finish 72-74% of them, but CUAs are closing the gap fast: 17% at the beginning of 2025, 45% today, and likely human parity in 2026.
Eventually, one will hit 100%. Then what?
A perfect score only proves a CUA can navigate any UI. It still won’t decide why a task matters, evaluate risk, or resolve ambiguity. The CUA execution layer is just the hands and eyes, not the brain.
Even with 100% capable CUAs, human workers will need to:
The value won’t be in the raw UI control by the CUA, it will be the scaffolding around it.
That said, once CUA execution is solved, every legacy desktop app will become an intelligent API surface. A typical use case will involve multiple models working together:
CUAs plus scaffolding map directly to Stage 4 (“AI uses your computer”), and then evolve into Stages 5 (“AI uses your computer without you”) and 6 (“multi-agent coordination”) in the 7‑stage human-AI collaboration roadmap.
Because each CUA will use existing workspaces and run with its own identity, existing IAM, DLP, session recording, and other guardrails will still apply as they do today. You won’t have to reinvent your security model.
Don’t get me wrong, it will be a big deal when CUAs hit 100 % execution. But this will only be a milestone, not the final destination, on the journey to AI in the workplace. When this happens, the value will shift to how well you design, secure, and govern the orchestration layer which drives the CUAs.
The scarcity (e.g. the value) will be the judgment around knowing what to do, how to do it, and what success looks like. When CUAs hit 100%, that judgement will still come from humans.
At that point, human workers will shift from doing to directing, and the org chart will start to look like a massive orchestration graph. Once CUA execution is solved, the only interface left to optimize will be your own thinking.
Read more & connect
Join the conversation and discuss this post on LinkedIn. You can find all my posts on my author page on the Citrix blog (or via RSS).
Video of my most recent talk
In May I gave the closing keynote at the EUCtech Denmark 2025 conference, called The Future of Work in an AI-Native World. I talked about a lot of what I covered today and walked through how AI will evolve and impact the workplace in the coming years. You can watch it on YouTube.
My upcoming talks
Brian Madden is a VP & futurist at Citrix. He writes about the future of work, AI in the workplace, and the evolution of Citrix.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。
