earakely-scale/cua-driver-bench

Cua Driver Bench for AgentEnv

trycua's Cua Driver Bench as an AgentEnv bundle that agent-env run runs straight from this repo. An agent shares a desktop with a person: in the background condition someone types in their own window the whole time, and the agent must do its task in the other windows without taking their focus, raising a window over theirs, moving their pointer, touching their clipboard or losing one of their keystrokes; in the foreground condition the desktop is the agent's alone. Either way it may use only the GUI.

All ten tasks of trycua's Bench v2, five runs each, in both conditions (MB-10 and MB-11 in the foreground only, as trycua's task manifest lists them: taskpack/v2-taskset.json, added in trycua/cua#6574 after the commit this bench pins): the multi-app workflows G02 to G04, and the BenchLab tasks MB-09 to MB-11 and IR-01 to IR-04. Each trial runs on a computer AgentEnv composes for it, Ubuntu 24.04 or macOS 26, with the task's apps, a computer-use driver and the bench's sentinel, the user's window; it is graded outside the computer by trycua's own evaluator, A22 criteria and GUI-only scan.

Our runs of it, with every recording, trajectory and grade, are in the dataset earakely-scale/CuaDriverBench-AgentEnv and its replay Space. These are not trycua's results.

Run it

On Linux, with Docker:

pip install 'agentenv-framework>=0.9.1446'
export LITELLM_BASE_URL=https://your-litellm-or-openai-compatible-proxy LITELLM_API_KEY=...
agent-env run hf://earakely-scale/cua-driver-bench --task cdb-g02-background --model anthropic/claude-sonnet-5-5   # one trial
agent-env run hf://earakely-scale/cua-driver-bench --eval bench-v2 --model anthropic/claude-sonnet-5-5             # every task of Bench v2 once
# GPT-6 in OpenAI's Codex CLI (agentenv-framework 0.9.1454 or later), on cua-driver's own tools as the board ran it
agent-env run hf://earakely-scale/cua-driver-bench --task codex-cdb-g02-background --model openai/gpt-6.1-sol --set envs/desktop-s02.tools=driver
# Vercel Labs' agent-desktop in cua-driver's place, as the Linux drivers comparison ran it (0.9.1456 or later)
agent-env run hf://earakely-scale/cua-driver-bench --task agent-desktop-cdb-g02-background --model anthropic/claude-sonnet-5-5

On an Apple silicon Mac, the same tasks run on macOS 26, in a logged-in session with Lume on the PATH and a Docker (colima will do). Select the Lume sandbox in a two-line config, which only a config can do, then run the macos- tasks:

mkdir -p .agentenv && printf '[sandbox]\ndefault = "lume"\n' > .agentenv/config.toml   # only a config picks the sandbox
agent-env run hf://earakely-scale/cua-driver-bench --task macos-ir-01-r1-foreground --model anthropic/claude-sonnet-5-5
agent-env run hf://earakely-scale/cua-driver-bench --eval macos-bench-v2 --model anthropic/claude-sonnet-5-5

The first run builds each task's computer, a few minutes on Linux and about 7 on macOS, where it first pulls Cua's macOS base (about 29 GB); later runs reuse it. On an amd64 Linux machine whose Docker can't emulate arm64, the build also says it "can't build linux/arm64" for each computer: that form is skipped, and the amd64 one runs. On macOS the drivers comparison's trials name their driver: macos-agent-desktop-… for Vercel Labs' agent-desktop, macos-macos-harness-… for browser-use's macOS harness, and macos-cua-driver-… for cua-driver on its own tools alone.

To run exactly what this card describes, pin the repo's commit: hf://earakely-scale/cua-driver-bench@<commit>. Each commit of this repo is the bench at one commit of its GitHub repo, where its tooling, tests and history live: this one is 1eb000c93cb32fd42389651a5a97ae3227f93ccd.

Run another driver or tool set

With agentenv-framework 0.9.1453 or later, --set envs/NAME.KEY=VALUE changes one of the bench's computer envs for a single run and leaves the bench as written. Each task deploys one computer env: desktop-benchlab for the precision and interruption tasks on Linux, desktop-s02, desktop-s03 and desktop-s04 for the multi-app tasks, and the same names ending in -macos on macOS. What a setting changes runs under ids ending in it, such as desktop-benchlab(desktop-benchlab.driver reference), so its trials stay apart from the bench's own.

pip install 'agentenv-framework>=0.9.1453'
# Linux: cua-driver on its own tools alone, as the dataset's macOS runs ran it
agent-env run hf://earakely-scale/cua-driver-bench --task ir-01-r1-background --set envs/desktop-benchlab.tools=driver
# Linux: AgentEnv's reference driver in cua-driver's place
agent-env run hf://earakely-scale/cua-driver-bench --task ir-01-r1-background --set envs/desktop-benchlab.driver=reference
# macOS: Vercel Labs' agent-desktop, or browser-use's macOS harness, on cua-driver's task and its own tools alone
agent-env run hf://earakely-scale/cua-driver-bench --task macos-ir-01-r1-background --set envs/desktop-benchlab-macos.driver=agent-desktop \
  --set envs/desktop-benchlab-macos.tools=driver
agent-env run hf://earakely-scale/cua-driver-bench --task macos-ir-01-r1-background --set envs/desktop-benchlab-macos.driver=macos-harness \
  --set envs/desktop-benchlab-macos.tools=driver

tools is standard (AgentEnv's computer_* tools alone), driver (the driver's own alone) or both, the default; the Linux runs in the CuaDriverBench-AgentEnv dataset ran with both. An eval deploys several envs, so set each one, for example --eval bench-v2 --set envs/desktop-benchlab.driver=reference --set envs/desktop-s02.driver=reference --set envs/desktop-s03.driver=reference --set envs/desktop-s04.driver=reference. The macos-agent-desktop-…, macos-macos-harness-… and macos-cua-driver-… task files are these settings written out, and they are what the dataset's macOS runs ran; in the foreground they also start agent-desktop headed, which --set can't. A driver that doesn't run on the computer, such as macos-harness on Linux, is refused before any task runs.

What's here

Path	What it is
bundle.toml	The bundle, and the agentenv-framework release it needs.
envs/	The computers (desktop*), the apps of each task, trycua's BenchLab and the bench's sentinel.
tasks/	One trial each: deploy the task's computer and the agent, prompt it in a condition, arm the sentinel, grade.
evals/	Sets of tasks: all of Bench v2 (bench-v2, macos-bench-v2), each condition, and smoke checks without a model.
artifacts/	The grader, the GUI-only check, the sentinel's steps, the checks, and trycua's task pack (artifacts/taskpack/).

artifacts/taskpack/ is trycua's task pack as published at trycua/cua@22a67eacd15dc71811e330ed88951dd615bb395d, each file as it is there, sha256 9fcd03b958b5c67611f604786d452ca27872c42643635722696b72e7dab82d11 over its MANIFEST.sha256, with trycua's licence beside it (artifacts/taskpack/LICENSE).

Licence

The tasks, their apps, prompts, evaluators and scoring are trycua's Cua Driver Bench, © 2026 Cua AI, Inc., under the Functional Source License 1.1, MIT Future License, redistributed here for non-commercial research, a Permitted Purpose under it. Everything else is the bench port's, under the Apache License 2.0. See LICENSE.md.