CUA-SWE

When Computer-Use Agents Meet Visual Software Engineering

Project website · Paper · PDF · GitHub · Interactive data browser · Leaderboard

105 tasks · 4 domains · Code editing + computer use · Deterministic verification

CUA-SWE evaluates whether agents can develop software by using the software they edit. Agents inspect running applications, discover visual requirements and runtime failures, change the source, and test their changes through computer use. Independent verifiers check the requested behavior and existing functionality.

Dataset contents

Configuration	Tasks	Applications
all (default)	105	All four domains
web	36	Editors, charts, maps, layout, and application workflows
game	29	Gameplay and interaction regressions across 12 games
mobile	20	Mobile-web applications with visual requirements and touch interactions
devops	20	Monitoring, dashboards, alerting, configuration, and observability

Each configuration contains a test split. The Parquet tables provide the task instructions and paths into the complete benchmark. The downloadable release includes all task source, assets, application environments, protected verifiers, reference/negative controls, evaluation pipelines, provider interfaces, documentation, and published results.

Load the tasks

pip install -U datasets huggingface_hub

from datasets import load_dataset

tasks = load_dataset("kingofspace0wzz/CUA-SWE", split="test")  # 105 tasks
web = load_dataset("kingofspace0wzz/CUA-SWE", "web", split="test")  # 36 tasks

print(web[0]["task_id"])
print(web[0]["instruction"])
print(web[0]["task_file"])

Choose game, mobile, or devops in place of web to load an individual domain. For a fixed dataset revision, pass the Hub commit hash as revision=.

Field	Meaning
task_id	Canonical benchmark task identifier
domain	web, game, mobile, or devops
instruction	Exact parsed task instruction; Mobile includes its instruction Markdown
task_file	Task specification path relative to the extracted release root
task_directory	Directory containing the complete task bundle
source_directory	Application source directory within the release
task_format	task_yaml or mobile_native_json
task_spec_sha256	SHA-256 of the published task specification bytes
instruction_sha256	SHA-256 of the instruction text encoded as UTF-8
benchmark_revision	Identifier of the packaged benchmark snapshot
archive_path	Complete benchmark archive in this dataset repository

Download the runnable benchmark

Download and extract the 358.4 MiB release archive:

hf download kingofspace0wzz/CUA-SWE releases/cua-swe-1.0.tar.gz \
  --repo-type dataset --local-dir cua-swe-download

tar -xzf cua-swe-download/releases/cua-swe-1.0.tar.gz
cd cua-swe-1.0

The archive includes the complete benchmark directory layout and its relative symlinks. Its SHA-256 is a4bfca7785c6962b50240baa7365b8beea76fef4eeea817ae413c640f13d6160. release.json and SHA256SUMS provide the release identity and download checksums.

To verify a complete Hub snapshot, run python tools/verify_release.py --package . from the dataset repository root. It checks all packaged files, every archive member, and all 105 task rows.

Run and evaluate agents

Use Linux and Python 3.11+. Web and DevOps use Node 22.22.2+; Game uses system Node 18 and Java 17. Install the sandbox and browser prerequisites in the environment setup guide.

From the extracted cua-swe-1.0 directory:

python3 -m venv "$HOME/.venvs/cua-swe"
source "$HOME/.venvs/cua-swe/bin/activate"
python -m pip install -e '.[evaluation]'
PLAYWRIGHT_SKIP_BROWSER_GC=1 python -m playwright install --with-deps chromium
npm install --prefix "$VIRTUAL_ENV/agent-tools" --install-strategy=nested @openai/codex@0.146.0
export PATH="$VIRTUAL_ENV/agent-tools/node_modules/.bin:$PATH"
python scripts/check_release.py

The baseline agent supports OpenAI Responses, OpenAI-compatible Chat Completions, Anthropic Messages, and Amazon Bedrock Converse. For example, configure a Chat Completions endpoint:

export MODEL_API_KEY="your-api-key"
export MODEL="your-model-id"
export API_BASE_URL="https://api.openai.com/v1"
API=(--provider chat-completions --model "$MODEL"
     --base-url "$API_BASE_URL" --api-key-env MODEL_API_KEY)

Alternatively, use a Bedrock API key with your region and model or inference-profile ID:

export AWS_REGION="your-aws-region"
export AWS_BEARER_TOKEN_BEDROCK="your-bedrock-api-key"
export MODEL="your-model-id-or-inference-profile-id"
API=(--provider bedrock --model "$MODEL" --region "$AWS_REGION")

See the provider guide for Responses, Anthropic, Bedrock Mantle, compatible services, and local servers. CUA mode uses models with image input and function calling.

Run one task first, then evaluate the complete domain:

python scripts/run_evaluation.py plan --domain web --scope full "${API[@]}"

python scripts/run_evaluation.py run --domain web "${API[@]}" \
  --task-id web.maplibre-reset-touch-reacquisition.001 --condition cua \
  --output-root "$PWD/artifacts/web-single"

python scripts/run_evaluation.py run --domain web --scope full "${API[@]}" \
  --condition cua --output-root "$PWD/artifacts/web-cua"
python scripts/run_evaluation.py run --domain game --scope full "${API[@]}" \
  --condition cua --output-root "$PWD/artifacts/game-cua"
python scripts/run_evaluation.py run --domain devops --scope full "${API[@]}" \
  --condition cua --output-root "$PWD/artifacts/devops-cua"

python scripts/configure_mobile_runtime.py \
  --output mobile-runtime-config.local.json --output-root "$PWD/artifacts/mobile"
python scripts/run_evaluation.py run --domain mobile --scope full "${API[@]}" \
  --condition cua --runtime-config mobile-runtime-config.local.json \
  --output-root "$PWD/artifacts/mobile/cua"

Use --condition code-only for source-and-terminal evaluation. The evaluator runs one attempt per selected task and writes patches, verification results, trajectories, and protocol evidence to the output directory. Report success separately for each domain and condition; evaluation documentation explains the result fields and treatment of infrastructure failures.

For your own agent loop or CLI, follow the agent integration guide. It specifies workspace access, tools, launch arguments, transcript format, and scoring handoff.

Leaderboard

Latest published results: September 24, 2026. Single-attempt task success (%), shown as code-only → CUA. Bold marks the best CUA score in each domain.

Model / agent	Web (36)	Game (29)	DevOps (20)	Mobile (20)
GPT-6 Astra (API)	27.8 → 66.7	17.2 → 37.9	0.0 → 80.0	0.0 → 55.0
GPT-5.6 Sol (API)	19.4 → 58.3	10.3 → 10.3	0.0 → 55.0	0.0 → 45.0
GPT-5.6 Luna (API)	11.1 → 5.6	0.0 → 6.9	0.0 → 15.0	0.0 → 35.0
GPT-5.6 Terra (API)	11.1 → 13.9	3.4 → 6.9	0.0 → 25.0	0.0 → 35.0
Claude Opus 5 (API)	22.2 → 55.6	13.8 → 20.7	0.0 → 60.0	deferred → deferred
Claude Opus 4.8 (API)	16.7 → 27.8	3.4 → 20.7	0.0 → 55.0	0.0 → 40.0
Claude Sonnet 5 (API)	8.3 → 0.0	6.9 → 3.4	0.0 → 40.0	0.0 → 35.0
Claude Fable 5 (API)	19.4 → 47.2	13.8 → 17.2	5.0 → 55.0	5.0 → 40.0
Grok 4.6 (API)	30.6 → 55.6	10.3 → 17.2	0.0 → 65.0	0.0 → 30.0
Codex + Sol	— → 55.6	— → 17.2	— → 55.0	— → 40.0
Claude Code + Opus 5	— → 55.6	— → 20.7	— → 65.0	— → deferred

— indicates no reported result; deferred indicates a postponed evaluation. Full results include repeated-attempt evaluations and scoring details.

Benchmark construction

The task inventory follows the canonical manifests: Web 36, Game 29, Mobile 20, and DevOps 20. Each task combines an application snapshot, a software-engineering instruction, and deterministic verification of the requested behavior. The paper describes task construction, control checks, and evaluation protocols.

Roadmap

• CUA-SWE 1.0: benchmark tasks, environments, domain evaluation pipelines, and baseline results.
• Multiple API interfaces: Responses, Chat Completions, Anthropic Messages, and Amazon Bedrock.
• More domain tasks: integrate additional applications, task families, and domains.
• Post-training RLVR environment: release environments and verifiable rewards for reinforcement learning with verifiable rewards.

Release and attribution

This dataset packages CUA-SWE 1.0. release.json identifies the packaged source snapshot and upstream provenance. The source-file inventory binds all archived files; the task-file inventory binds the canonical task inputs. Published model scores retain their original evaluation date.

License and attribution files are included with the task snapshots. See third-party notices and release licensing for the applicable terms.

Citation

@article{wang2026cuaswe,
  title={CUA-SWE: When Computer-Use Agents Meet Visual Software Engineering},
  author={Wang, Prince Zizhuang and Liang, Chenhao and Xu, Zelong and Yuan, Aojie and Zhou, Xiaolin and Zhang, Haiyue and Zhao, Yue and Hu, Xiyang and Jiang, Shuli},
  journal={arXiv preprint arXiv:2609.32600},
  year={2026}
}