agent-desktop completed 26 of 35 tasks on a Mac, cua 33, and Codex 34

@argofowl published a short comparison on X of three tools that control a computer in place of a human: agent-desktop, cua and Codex's computer use. He ran all three on his own Mac and gave each the same set of 35 tasks. The results: agent-desktop on Opus 5.5 completed 26 tasks, cua on the same Opus 5.5 completed 33, and Codex on Terra 6.1 completed 34. It is a post of a few lines, not a study, but the numbers are concrete enough to pick apart.

Three tools and their results

agent-desktop (Opus 5.5) completed 26 of 35 tasks, about 74%. Nine tasks went unfinished. The author names three types of failure: dragging (drag), clicks on a canvas and scrolling a huge list. The end of the sentence in the post is cut off, so there is nothing to say about the causes. Judging by context, it is agent-desktop that is meant, but the text itself does not spell this out.

cua (Opus 5.5) completed 33 of 35 tasks, about 94%. It missed two tasks. The author does not say which ones.

Codex's computer use (Terra 6.1) posted the best result: 34 of 35 tasks, about 97%. One task was left undone, and it, too, is not described.

The gap between first and last place is 8 tasks. Between agent-desktop and cua it is 7 tasks, with the same Opus 5.5 model. This is the most telling part of the comparison. It shows that the outcome depends not only on the model but also on the layer around it. That layer determines how the agent sees the screen, how it forms actions and how it passes them to the system.

Which tasks proved hardest

The three task types on which agent-desktop failed are quite typical for agents that work with a graphical interface. Dragging requires a precise sequence: press, move, release. Any delay or inaccurate coordinate breaks the action. Clicking on a canvas is hard because a canvas has none of the usual interface elements with names and roles. The agent has to rely on pixels. Scrolling a huge list is a test of endurance: the agent has to keep scrolling without losing its place and without going in circles.

It is worth stopping here and saying it plainly: the post does not explain why agent-desktop in particular failed. The cause could be limits of the tool itself, its input method, its screenshot handling or its settings. The idea that drag and canvas are its weak spot can be treated as a hypothesis, not an established fact.

What can and cannot be drawn from this

There is too little data for strong conclusions. First, the sample: 35 tasks and one computer. One extra miss shifts the result by almost 3 percentage points. So the difference between 33 and 34 tasks for cua and Codex is within the noise, and Codex cannot be called the outright leader. Second, it is unknown how many times each configuration was run. Third, the makeup of the tasks is not disclosed, which means the test cannot be reproduced. Fourth, Codex uses a different model, Terra 6.1, so its comparison with cua and agent-desktop does not isolate the effect of the tool. A clean harness-to-harness comparison exists only between agent-desktop and cua: they share Opus 5.5.

The practical takeaway for anyone choosing a UI automation tool is fairly down-to-earth. If your scenarios include dragging, canvas work or long lists, test exactly those before rollout. General claims that an agent "can control a computer" usually do not cover such cases. Your own set of 30–50 typical tasks, run on your own working machine, will tell you more than someone else's results. @argofowl's post is useful as an example of that approach and as a signal that the spread between harnesses on a single model can reach seven tasks out of thirty-five.

The post drew about 5 thousand views, so it remained a note for a narrow circle. If the author publishes the task list and details of the failures, the comparison will become considerably more valuable.