Cua Driver Bench on AgentEnv

Cua Driver Bench on AgentEnv

A replay viewer for computer-use agents running Cua Driver Bench tasks, with the recorded desktop, tool calls, grades and drivers side by side. These are not the benchmark makers' official results.

Made by earakely-scale on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Cua Driver Bench on AgentEnv is a replay viewer for computer-use agents. Each replay shows the recorded desktop, the tool calls the agent made and the grade. In the background condition it also shows a simulated user typing in a separate window. When more than one driver ran on identical tasks and one model, you can line the drivers up next to each other.

Who it suits: People who evaluate or build computer-use agents will get the most from it. The tasks come from trycua's Cua Driver Bench and ran on AgentEnv computers. The rows below are judgment from what the page shows.

Industry | Who uses it | How they use it | Example
Research | Agent evaluators | Compare how different drivers handle the same task with the same model | Watching two drivers on one task side by side
Software | Computer-use agent developers | Replay a trial to see each tool call and where the agent went wrong | Stepping through a failed trial and reading its grade
Product | Teams choosing a driver | Look at recorded runs and grades when picking a driver | Checking how drivers behave while a user types nearby
Education | Students of agent evaluation | See what a benchmark trial looks like from desktop to grade | Replaying one trial from start to finish

You can skip it if you want to run the benchmark yourself, since this page only replays recorded trials.

Where to use it: It runs in a browser tab, and its files open locally on any desktop that has a browser.

• A web browser: open the Space page on Hugging Face
• Windows, macOS or Linux: download the files and open index.html in a browser
• A Hugging Face dataset: the trials, recordings and grades are kept there and streamed to the page

Use case | Who | What they get back
Replay a trial | Agent developers | The recorded desktop, the agent's tool calls and the grade
Compare drivers | Evaluators | Several drivers on the same tasks and model, side by side
Check the background condition | Researchers | The simulated user's typing in its own window, next to the agent's work

How to set it up: Using the page online needs nothing. A local copy needs a browser and Python with pip. The page streams each trial from the dataset, so a local copy may still need a network connection.

• A web browser
• Python with pip, only for downloading a copy
• Time for the download, since the repository holds 4985 files

1. Install the Hugging Face hub package with pip.
2. Download the Space's files into a folder.
3. Move into that folder.
4. Open index.html in a web browser.

pip install -U huggingface_hub
hf download earakely-scale/CuaDriverBench-AgentEnv --repo-type space --local-dir CuaDriverBench-AgentEnv

How to use it: Open the Space page and pick a trial to replay. The README gives no click-by-click guide, so follow what the page offers. With a local copy, go into the folder and open index.html.

cd CuaDriverBench-AgentEnv

Tool | What it does
Desktop replay | Plays the desktop as it was recorded
Tool calls | Lists the calls the agent made during the trial
Background user | Shows the simulated user typing in its own window
Grade | Gives the score the trial received
Driver comparison | Puts drivers that ran the same tasks and model side by side

Example: Open the page, pick one trial and play its replay. You see the desktop as recorded, each tool call the agent made and the final grade. If several drivers ran that task, a side-by-side view lets you compare them.

For newcomers and experts: A computer-use agent is an AI that works a computer by clicking and typing, much as you do. Start by replaying one trial from beginning to end, and read its grade last.

The page is static HTML plus app.js, with a data folder in the repository. The dataset holds the trials, recordings and grades, and the page streams each trial's web copy from there. The tasks, apps, prompts and evaluators belong to trycua, so the source states plainly that these results are not official.

A trial is recorded on an AgentEnv computer
Its recording and grade go into the dataset
The page streams the trial from the dataset
You replay it or compare drivers

Good to know:

• License: Apache-2.0 for the Space's own code, while trycua's tasks, apps, prompts and evaluators keep their own terms from Cua AI
• Cost to run: none named
• Data it reaches: the dataset on Hugging Face that the page streams each trial from
• Fonts: Inter and JetBrains Mono, both under the SIL Open Font License
• Not for: official benchmark results, since the source says these are not trycua's official results

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: static
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/earakely-scale/CuaDriverBench-AgentEnv
• Files: https://huggingface.co/spaces/earakely-scale/CuaDriverBench-AgentEnv/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.