Most of the software a small business runs was built for a person at a screen: the supplier portal, the old accounting package, the government form. An AI agent that could work those screens the way an employee does, looking and clicking, would reach all of it without waiting for anyone to build an integration. The Qwen team at Alibaba, with the Xlang Lab, has published a detailed account of what that kind of agent takes: a technical report on Qwen-CUA, an agent that sees only screenshots and acts only through the mouse and keyboard.[1] The headline benchmark numbers are strong. The more useful result sits near the end: when the team gave the agent a faster tool on top of the screen, it finished fewer tasks.
An agent that works the screen like an employee.
"Native" computer use, in the report's strict sense, means the agent gets what a person at the desk gets and nothing more: a screenshot in, a mouse or keyboard action out. No page code, no accessibility tree, no shell, no task-specific API.[1] The appeal for a business is reach. Anything a person can operate on a screen, the agent can at least attempt, including software that will never get an integration.
Under the hood is a 397-billion-parameter mixture-of-experts model (about 17 billion parameters active at a time), trained with the help of roughly 40,000 tasks built so their outcome can be checked automatically, on a fleet with access to nearly 100,000 virtual CPUs.[1] On OSWorld-Verified, a standard test of everyday desktop tasks, the team reports 86.2, against 83.4 for Claude Opus 4.8 and 78.7 for GPT-5.5 in the same table. Across the seven capability tests in that table, Qwen-CUA scores highest on two (OSWorld-Verified and MacAgentBench), Claude Opus 4.8 on four (including MyPCBench, the personal-desktop test used below), and GPT-5.5 on one.[1] Read the leaderboard as one vendor's report on tests it selected. The scale is the other lesson: this is not a weekend build, and, as the rest of this note shows, the people who built it say it belongs inside a hybrid agent.
The faster tool that made it worse.
MyPCBench drops an agent into a simulated personal computer: an Ubuntu desktop with 17 web applications, an office suite, the user's files and a history across them, 184 tasks in all.[1] The team ran the same tasks two ways: screen-only, and with a Bash tool added so the agent could type commands straight into the operating system when that was quicker than clicking.[1]
In the Qwen team's own run of the MyPCBench test, adding a Bash command-line tool cut Qwen-CUA's average turns per task from 63.6 to 49.1 and lowered its task score from 58.7% to 55.1%. The same happened to Qwen3.7, the Qwen model the report uses as its baseline: turns fell from 69.3 to 53.4, and the score from 51.6% to 41.8%. The report calls that score "task performance"; its screen-only values match the strict rate it defines elsewhere, the share of tasks finished perfectly with every grading criterion met, so that is how we read it.[1]
The team's explanation is specific: its models "have not yet fully learned when to switch between native computer use and command-line execution," and "unnecessary or poorly timed switches can reduce task completion."[1] They call it an optimization gap rather than a flaw in the hybrid idea, and see room for training aimed at tool choice to recover the loss while keeping the shorter paths.[1] That may happen. What they measured, today, on their own models, is that a faster tool without the judgment to use it cost completed work.
Why it matters before you automate.
When an agent struggles, the instinct is to give it more: another integration, another tool, a faster path. This result says each tool also adds a choice the agent has to get right at every step, and a wrong choice can cost more than the tool saves. Two models on one benchmark is thin evidence, but it matches the advice we already give: start with the simplest thing that does the job and add capability one piece at a time, measuring as you go. Our note on the order autonomy is earned in walks through it. Three habits of ours follow from it; the report does not propose them.
- Add one tool at a time, and rerun the same tasks.
Keep a fixed set of real tasks from your own work. Run it before and after each new tool or integration. If the agent got faster but finished less, the tool is not ready, or the agent does not yet know when to reach for it.
→ the test set stays the same; only the tool changes - Count finished tasks, not steps.
Shorter runs look like progress on a dashboard. In this report, shorter runs came with fewer completions. Track how many tasks reach a correct end state, and treat speed as the second number.
- Write down when to use each tool.
An agent picks tools partly from their descriptions, so state the cases each one covers and the cases it does not: "use the command line to move or rename files; use the screen for anything inside the accounting application." That is the routing the Qwen team says its models had not learned, and you can write part of it down. The description is one of the four parts of a tool in our WebMCP note.
Pixels are the fallback, not the fast path.
The report's own conclusion is the sentence most worth a business owner's attention: "Generality, however, does not imply optimality."[1] Reading the screen, reasoning, and making one click at a time is slow and costly, and the team says plainly that visual operation "is also inefficient when a task can be expressed directly through code, shell commands, or structured APIs." Its recommendation is a hybrid agent that uses code for precise, high-volume work, structured tools when they exist, and the screen "whenever a visual or closed interface must be operated." The team calls native computer use "the universal grounding and fallback layer of a hybrid agent," not the only interface.[1]
That is the builder of a screen-driving agent saying not to use one for everything. For an owner deciding what to automate, it sorts the work: if a system has an API or an export, use it, and keep the screen-driving agent for the portals and older applications that have neither. Our WebMCP note covers the other half of the same decision: websites that declare tools so agents do not have to drive the screen at all.
The report also makes an efficiency case for its own model, and it needs a careful read. On OSWorld-Verified, Qwen-CUA reached 86.2 using about 3,606 output tokens per task, the text the model writes while it reasons and acts, which the report says mainly drives cost and waiting time. The team reports that Claude Opus 4.8 reached 80.0 "at a similar budget" and 83.3 when it used 21.8K tokens (83.4 in the report's main table, 83.3 in its efficiency passage).[1] What that comparison supports is narrow. The report's own conclusion from it is that Qwen-CUA's gain "does not simply result from more verbose reasoning." It counts output tokens only; the screenshots the agent reads every turn are not in it, so it is not a cost-per-task comparison.
Turn counts are where to slow down. On the harder OSWorld 2.0, Qwen-CUA averaged 218.9 turns per task, against 83.5 for GPT-5.5 and 105.7 for Opus 4.8, and the team itself warns this is not a like-for-like measure: the two proprietary agents can batch several actions into one turn, while Qwen-CUA makes one action per turn.[1] Neither "slower" nor "faster" follows from those numbers. If anyone sells you a turn count as a speed claim, ask what a turn contains.
Which agent you deploy matters.
An agent that reads screens reads whatever is on them, including text someone planted there for it. Indirect prompt injection is the name for that: instructions hidden in a web page, a shared file or a chat message, which the agent mistakes for its orders. The report tests it with RedTeamCUA, a benchmark that plants such instructions in tasks across a file-sharing application (ownCloud), a team chat (Rocket.Chat) and a forum (Reddit), and records two numbers: how often the agent still finishes the real task, and how often the attack succeeds.[1]
Qwen-CUA cut the attack success rate from 36.6% for the Qwen3.7 baseline to 16.4%, a real improvement, and close to GPT-5.5 at 15.6%. In the same table, the vendor's own, Claude Opus 4.8 sits at 0.7% while also finishing the most real tasks, 80.7%.[1] The team's own reading is that Opus 4.8 "remains strongest on both metrics," that its own 16.4% "indicates residual risk," and that the benchmark shows improved resistance, "not a deployment-safety guarantee."[1]
Two cautions keep the table honest. The chat platform is hard for every agent (Qwen-CUA finished 32.6% of those tasks), and the team notes that an agent which fails to act also fails to follow the planted instructions, so a low attack rate there is partly failure, not resistance.[1] And this is one benchmark, in one vendor's report.
For a business, the useful lesson is about the checklist. Resistance to planted instructions can differ by more than an order of magnitude between agents that look close on capability tests, so it belongs on the checklist next to accuracy and price. Ask a vendor for its injection-test results, with the benchmark and the date. Keep consequential actions behind a person: in the report's own browser showcase, the agent asks the user to confirm before consequential operations.[1] Our notes on setup prompts written for your AI and agents that infect each other cover the same exposure from other angles.
What this does not claim.
Every figure is the vendor's. The Qwen team built the model, chose the benchmarks and compiled the comparison; by its own account, most competitor scores come from official reports, not its own runs. Nothing here has been independently replicated.
The Bash result is narrow. Two Qwen models, one benchmark, one way of adding the tool, and authors who see room for training to close the gap. It shows that a faster tool can cost completions, not that it always will. The report gives no run-to-run spread, so the 3.6-point drop for Qwen-CUA shows a direction; its size is unsettled.
Turn counts do not compare across vendors. The proprietary agents batch actions per turn and Qwen-CUA does not. The report says so itself.
The safety table is not a safety rating. RedTeamCUA measures one kind of attack on three platforms. The authors call the result "not a deployment-safety guarantee," and low attack rates on the chat platform partly reflect failed tasks.
No client record. We have not deployed Qwen-CUA, and the report describes its browser deployment as internal without saying whether the agent is available to outside users. The three habits above are our practice, not findings from the report.
Wondering whether an agent could take over the work in a supplier portal or an older desktop application? Describe the system through the contact form and you get a free written first read: whether there is an API path first, and where a screen-driving agent would fit. Agent work like this falls under our assistants and agents services.