Skip to content

AgentOS is an open source python API and a command line interface for building, running, and sharing learning agents.

License

Notifications You must be signed in to change notification settings

andyk/agentos

 
 

Repository files navigation

The Python Component System and AgentOS

This project consists of two major pieces: the Python Component System (PCS) and AgentOS.

Python Component System (PCS)

PCS is an open source Python API, command line interface, and public web registry for building, running, and sharing Python programs. The goals of PCS are to:

  • Make Python program execution reproducible.
  • Transparently manage Python virtual environments while providing a Python API for pip and virtualenv.
  • Simplify experiment tracking and code sharing.

PCS does this by allowing you to explicitly specify dependencies and arguments for your program and then providing a thin runtime (currently based on MLflow) to automatically instrument your program's execution. PCS is compatible with most frameworks that are used to build machine learning and reinforcement learning systems.

AgentOS

AgentOS is a set of libraries built on top of the Python Component System that make it easy to build, run, and share agents that use Reinforcement Learning (RL) to solve tasks.

Key features of AgentOS:

  • Easy to use Agent API for developing and running new agents.
  • A public repository of popular RL environments and agents, and runs of those agents in those environments that can be reproduced with a single line of code.
  • Example learning agents from different disciplines and research areas are available in the example_agents directory of the project source code.

Connect

Ask questions or report bugs in PCS and AgentOS in GitHub Issues or on the dev Discord.

Find the AgentOS source code on Github.

Test Status Indicator

The Python Component System and AgentOS are alpha software; APIs and overall architecture are likely to change significantly over time. They are licensed under the Apache License, Version 2.0.

Quickstart

See the agentos.org quickstarts.

Documentation

For detailed documentation see the agentos.org docs.

Development Process

AgentOS uses GitHub Issues to track development work. Submit any bug reports or feature requests to this issues tracker.

For significant feature work (more than a couple dev days or something that fundamentally changes internal or external interfaces), we run a design process to solicit feedback from other collaborators. Read more about this process in the Proposing Features section.

To contribute to AgentOS, the general workflow is as follows:

  • Sync with the core development team via the issue tracker so we can avoid unnecessary or duplicated work.
  • Fork the AgentOS repo.
  • Complete your feature work on a branch in your forked repo. Ensure all checks and tests pass.
  • Issue a pull request from your forked repo into the central AgentOS repo. Assign a core developer to review.
  • Address any comments and the core developer will merge once the PR looks good.

Proposing Features

For new features and other big chunks of work, AgentOS uses a design process centered around design proposals, discussions, and design docs. The goal of the process is to:

  • Allow developers to think through a design, and
  • Allow stakeholders to give feedback

...before development begins.

If you'd like to propose a feature, please follow the procedure found in the design_docs README. You can also browse existing design docs in the folder to get a feel for the general content and style.

Installing AgentOS From Source

To install agentos from source (e.g., to play with the example_agents), run the following:

git clone https://github.com/agentos-project/agentos.git
pip install -e agentos # you may want to do this in a virtualenv or conda env.

Testing

To run tests, first install the requirements (note, this script installs the Python requirements into the currently active virtual environment):

cd agentos # the project root, not the nested agentos/agentos dir
python install_requirements.py

Then run the tests:

pytest

Also, we use Github Actions to run tests with every commit and pull request (see the test workflow)

If you want to the CLI to interact with a local development server, define the environment variable (or create a .env file) USE_LOCAL_SERVER=True.

To run website tests:

python install_requirements.py
cd web # the web directory contained in project root
python manage.py test

Note that some tests (e.g., see web/registry/tests/test_integration.py) test functionality for interacting with github repositories by fetching code from https://github.com/agentos-project/agentos. Where possible, in order to make it easy to have those tests run against code in a github repo that you can change during development without disrupting other PRs, the test code uses global variables defined in tests/utils.py to decide which github repo to use when testing.

If you make changes to code that is fetched from github for use by tests, then please follow this process for your PR:

  1. While doing development, change the TESTING_GITHUB_REPO_URL and/or TESTING_BRANCH_NAME global variables in tests/utils.py to point to a version of your PR branch that you've pushed to github. We recommend commenting out the default "prod" values of these variables so that you can uncomment them in the next step when the PR is approved for merge.
  2. After your PR is approved and right before it is merged, push the branch you used during testing to the test_prod branch of the agentos-project account https://github.com/agentos-project/agentos.git. And then update the variables in tests/utils.py (you should be able to just uncomment the lines you commented out in step 1 above, and delete the lines you added).

Building Docs

The documentation source is in the documentation directory and written in ReStructuredText. The docs are built using Sphinx. To build the docs, first install the dev requirements (note, this script will install requirements into the currently active Python virtual environment):

python install_requirements.py

Then use the build script:

python scripts/build_docs.py

Use the --help flag to learn more about other optional flags that build_docs.py takes, including --release (for publishing the docs) and --watch (for auto-recompiling the docs whenever doc source files are changed).

Notice that the build file puts the compiled docs into docs/<version_num> where version_num comes from agentos/version.py.

Or you can build the docs manually (e.g., to control where output goes):

sphinx-build documentation outdir  # Or use sphinx-autobuild.
# Open and inspect outdir/index.html in your browser.

Publishing Docs to agentos.org

agentos.org is a github.io website where the AgentOS docs are hosted. To publish updated docs to agentos.org, checkout the website branch and build the docs per the instructions above, then create a PR against the agentos-dev/website branch. Once committed, those changes will become live at agentos.org automatically.

Assuming you have local branches tracking both the master and website branches, and all changes to the documentation source files have all been committed in the master branch, the workflow to publish updated docs to agentos.org might look similar to:

git checkout website
git merge master
python scripts/build_docs.py --release -a  # The -a is a `sphinx-build` flag.
git add docs
git commit -m "push updated docs to website for version X.Y.Z"
git push

Building README.rst

The main project README.rst is built via the script python scripts/build_readme.py, which re-uses sections of documentation. This avoids duplication of efforts and lowers the chances that a developer will forget to update one or the either of the README or the docs.

To update README.rst, first familiarize yourself with its build script scripts/build_readme.py. There you can see which sections of documentation are included in README.rst, plus some text that is manually inserted directly into README.rst (e.g., the footer).

Releasing

Here are the steps for releasing AgentOS:

  1. Build and check the distribution artifacts for the release by running:

    python install_requirements.py
    python setup.py sdist --formats=gztar,zip bdist_wheel
    twine check dist/*
    

    This will create a wheel file as well as tar.gz and zip source distribution files, and catch any blockers that PyPI would raise at upload time. Fix any errors before proceeding.

  2. Create a release pull request (PR) that:

    • Removes "-alpha" suffix from the version number in agentos/version.py.
    • Contains draft release notes (summary of major changes).
  3. Wait till the PR gets LGTMs from all other committers, then merge it.

  4. Build and publish the docs for the new version, which involves creating a pull request against website branch. This is required for all releases, even if the docs have not changed, since the docs are versioned. When you run the build_docs.py script, you will use the --release flag (see Building Docs & Publishing Docs to agentos.org for more details).

  5. Create another follow-on PR that bumps version number to be X.Y.Z-alpha which reflects that work going forward will be part of the next release (we use semantic versioning).

  6. Push the release to PyPI (see Pushing Releases to PyPI).

  7. Create a github release and upload the tar.gz and zip source code distribution files. This will create a git tag. For the tag name, use "vX.Y.Z" (e.g. v0.1.0).

Pushing Releases to PyPI

We make AgentOS available in PyPI. To push a release to PyPI, you can approximately follow these python.org instructions, which will probably look something like:

python install_requirements.py
rm -rf dist
python setup.py sdist --formats=gztar bdist_wheel
twine check dist/*
twine upload dist/*

This README was compiled from the project documentation via: python scripts/build_readme.py.

About

AgentOS is an open source python API and a command line interface for building, running, and sharing learning agents.

Resources

License

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 95.5%
  • HTML 3.7%
  • Other 0.8%