py-agent

A monorepo for two related Python packages:

• py-agent-lib — minimal, domain-agnostic DAG scheduler/executor for step-based workflows. Pydantic v2 throughout. The runtime substrate.
• py-agent-ner — NER training-data generation built on py-agent-lib. Bring your own labels, LLM, and downstream trainer.

Both publish independently to PyPI; both share this repository.

Layout

py-agent/
├── packages/
│   ├── py-agent-lib/        # the runtime: types, plan, executor, adapters
│   └── py-agent-ner/        # the NER toolkit: models, pipeline, BIO conversion
├── examples/
│   └── quickstart/          # end-to-end NER demo (gemma4 via Ollama)
├── pyproject.toml           # uv workspace root (not published)
└── LICENSE                  # MIT, covers both packages

examples/ lives at the repo root, not inside any package, because it demonstrates the cross-package integration — you wouldn't ship an example inside the library it imports from.

Develop

git clone https://github.com/gaslit-ai/py-agent.git
cd py-agent
uv sync                       # installs both packages editable in one .venv
uv run pytest                 # runs all tests across both packages
uv run python examples/quickstart/run.py   # live demo (needs Ollama + gemma4)

Within the workspace, py-agent-ner's dependency on py-agent-lib resolves to the in-tree package, so you can change py-agent-lib's API and immediately see how py-agent-ner reacts in tests.

Publish

Each package is published independently from this repo. Tag with a package-name prefix:

# When py-agent-lib is ready for 0.1.0:
git tag lib-v0.1.0 && git push --tags

# When py-agent-ner is ready for 0.1.0:
git tag ner-v0.1.0 && git push --tags

The matching workflow under .github/workflows/ builds and publishes the right package. Each package has its own version number — they evolve independently.

Quick links

• Library API → packages/py-agent-lib/README.md
• NER toolkit API → packages/py-agent-ner/README.md
• Working demo → examples/quickstart/README.md