USEagent: Unified Software Engineering Agent

USEagent is the first unified software engineering agent built with a top-down abstraction. USEagent abstracts "unit of work" in software engineering into an individual action, based on software engineering practices. Actions can be maintained and optimized separately. USEagent handles different types of software engineering tasks by dynamically orchestrating the actions.

Getting started

Build USEagent with docker

It is strongly recommended to run USEagent with docker containers. The preferred mode is to to start a USEagent container and mount relevant files in as a copy.

Build docker image with just the agent:

DOCKER_BUILDKIT=1 docker build --build-arg USEBENCH_ENABLED=false -t useagent:dev .

Build image on top of existing image (useful when running on benchmarks, example script assumes you have a my.project.to_work_on image):

DOCKER_BUILDKIT=1 docker build --build-arg BASE_IMAGE=my.project.to_work_on --ssh default -t useagent:dev .

For working on USEbench, see notes on usebench.

Build USEagent without docker

If you are in an isolated/containerized environment, you may want to use USEagent directly:

Install UV

uv sync
uv build
uv run <<args>>

Run USEagent on local tasks

Example run in a (fresh) local folder:

mkdir ./useagent-tmp
echo "Hi, this is a simple file" > ./useagent-tmp/README.md

docker run --rm \
  --name useagent-test \
  -e GEMINI_API_KEY=YOUR_KEY_GOES_HERE \
  -v ./useagent-tmp:/input \
  useagent:dev \
  useagent local --model google-gla:gemini-2.0-flash --task-description 'write a shell file that prints a vegan tiramisu recipe' --output-dir /output --project-directory /input"

Inside the docker container, there will be a working-copy of your project created, so any files you mount in are safe and will remain unchanged.

Run USEagent on Github Repository

Example will checkout the repository:

docker run --rm \
  --name useagent-test \
  -e GEMINI_API_KEY=YOUR_KEY_GOES_HERE \
  -v ./useagent-tmp-out:/output \
  useagent:dev \
  useagent github --model google-gla:gemini-2.0-flash --task-description 'write a shell file that prints a vegan tiramisu recipe' --repo-url https://github.com/octocat/Hello-World.git --output-dir /output"

Run USEagent on other examples

There are more example scripts as well as scripts to run on benchmarks.

They assume that the correct key (see below) are in a .env file and available in the runtime.

Configuration

Supported --models:

Model	Required ENV
'google-gla:gemini-2.0-flash'	GEMINI_API_KEY
'openai:gpt-4o'	OPENAI_API_KEY
'anthropic:claude-3-5-sonnet-latest'	ANTHROPIC_API_KEY

See more at pydantic API Documentation

License / Terms of Agreement

Code within this project follows the Apache 2.0 License.

We further provide a local copy of the gemma-3-4b-it model, which comes with terms of usage. Any usage or distribution must comply with them.

Citation & Contact

A preprint is available on arxiv.

To cite please use:

@inproceedings{applis2026unified,
  title={Unified Software Engineering Agent as AI Software Engineer},
  author={Applis, Leonhard and Zhang, Yuntong and Liang, Shanchao and Jiang, Nan and Tan, Lin and Roychoudhury, Abhik},
  booktitle={2026 IEEE/ACM 48th International Conference on Software Engineering (ICSE)},
  year={2026}
}

For queries, you can open an issue, or contact Yuntong Zhang (Email: [email protected]).