Marvin

Marvin, from the team behind Prefect, is a Python framework that treats LLMs as a runtime for ordinary code. Its signature idea is the AI function: you declare a function signature and a docstring describing what it should do, and Marvin generates the output on demand — no prompt templates, no parsing logic. Around that core it layers discrete observable tasks, specialized agents assigned to each task, and threads that orchestrate tasks into bigger behaviors.

Version 3.0 rebuilt it on Pydantic AI model support (OpenAI by default, any Pydantic AI provider works), with casting, classification, extraction, and generation as first-class primitives. At 6,200 stars it is a mid-size player — respected, well-documented, and honest about its scope: it is the 'keep it simple' counterweight to heavier agent frameworks. The 3.0 rewrite means migration churn if you used the 2.x API, and the ecosystem of examples is thinner than LangChain's.

1. Install: uv add marvin (or pip install marvin), Python 3.10+.
2. Export your provider key: export OPENAI_API_KEY=... (any Pydantic AI model works).
3. Define an AI function with a signature and docstring, or a Task with assigned agents.
4. Compose tasks into a Thread for multi-step orchestration.
5. Run, inspect the observable task traces, and iterate on the docstrings — the docstrings are the program.