Cua launches Driver Bench for desktop agents without user interruptions

• Background macOS full-pass rates fell to 58% for Cua Driver, 33% for Codex computer use, 20% for agent-desktop and 3% for browser-use in Cua's launch results.
• Completing the task is only one requirement: the scoring rules also require GUI-only execution and zero disturbance across eight user-interruption checks.
• Cua's macOS lead over Codex remains within the error bars, as Cua acknowledges.
• Windows topped out at 28%; both tested Linux drivers reached 13%, according to the platform results.

Claude Sonnet 5.5 makes the decisions even for OpenAI's Codex computer-use tool layer. A 50 ms pop-up menu can spoil an otherwise clean run, while two hover-dependent tasks are excluded from the background score entirely.

Every driver gets the same model, prompts and budget: 360 seconds and 45 turns, according to Cua's report. A simulated user types one character per second in a foreground window while the agent works behind it.

A full background pass requires:

1. Task completed. A checker reads the app's final state; no model grades the result.
2. GUI-only execution. A guard rejects shell or filesystem shortcuts through terminal apps or driver tools.
3. User undisturbed. None of these eight events may occur:

An unmeasured disturbance criterion counts as a failure. Driver-owned overlays, such as an agent cursor, are logged but exempted.

Cua reports these macOS results, combining Scale's AgentEnv runs with its own Codex run:

| Driver | Foreground full pass | Background full pass | Run by |
| --- | ---: | ---: | --- |
| Cua Driver | 78% (39/50) | 58% (23/40) | Scale |
| Codex computer use | 53% (16/30) | 33% (13/40) | Cua |
| Vercel agent-desktop | 70% (35/50) | 20% (8/40) | Scale |
| browser-use macos-harness | 64% (32/50) | 3% (1/40) | Scale |

Foreground testing covers ten tasks; background testing covers eight. Those different task sets prevent a strictly like-for-like foreground-to-background comparison.

The macOS lead is directional, Cua cautions: its 95% Wilson interval is 42–71%, overlapping Codex's 20–48%.

Browser-use completed 32 of 40 background tasks within the GUI-only rule, but disturbed the user in 31 of those completions, according to the completion breakdown. Task-completion scores alone flatter it badly here.

The GUI-only guard caught concrete shortcuts in Cua's detailed results:

• Codex on macOS: six of 15 multi-app trials used the plugin's JavaScript tool to access Node's fs or child_process modules. Those trials failed.
• browser-use on Windows: 27 of 63 scored trials violated the rule after agents requested PowerShell, cmd or subprocess execution.

Cua also changed a Windows rule during the November 6 run. Its methodology exempts shells launched directly by a driver's own binary as tool internals; agent-supplied code that launches a shell still fails.

Cua Driver itself uses PowerShell to inspect browser processes. Seven CDB-G04 trials stopped under the original rule were rerun with the same seeds; Cua says applying the revision to earlier Windows results changed no other driver's verdicts.

All Windows results came from one Windows 11 VM run by Cua. No driver achieved a full background pass on any of the three multi-app tasks, leaving each at 0/15, according to the Windows breakdown.

Codex's zero reflects its foreground-only Windows design under a benchmark that requires background operation.

Both Linux drivers' five passes came from IR-02, the banner task. In the XFCE runs, xfwm4 gave new modal dialogs keyboard focus, causing all 15 dialog trials per driver to disturb the user despite completing the task.

Cua documents three remaining input problems in its platform analysis:

• Pop-up timing: Scale's checker caught the roughly 50 ms macOS menu flash in one of five trials; Cua's checker caught it in four. Cua attributes the difference to Scale sampling at 20 Hz.
• Dragging: Cua Driver passed zero of five background drag trials on macOS, versus five of five on Windows.
• Hover: no tested platform provides reliable, supported background hover without moving the user's real pointer.

Cua asks OS vendors for three primitives:

1. Window-targeted input without foreground activation.
2. An independent pointer and hover state.
3. Dialog responses delivered in place.

Scale publishes per-trial recordings and agent trajectories, with replays pairing tool calls and simulated user typing. Reproduction resources are linked from Cua's release thread.

Cua wrote the tasks and scorer. Its method summary acknowledges that Cua Driver fixes were developed against these tasks; a sealed held-out task set is planned to measure how much that tuning matters.