2026-07-17 –, S3B
How to design a framework for building agents without it being a framework? What if building agents didn't require complex abstractions and APIs?
We are presenting the new Python AI SDK - a toolkit built on just a few powerful, fundamental ideas:
You don't need abstractions, you need composable primitives. You don't need graphs, your Python code is already a graph. With the field evolving this fast, less is more.
State is king. Keeping the state minimal, safe, and serializable unlocks long-running, durable agents.
Everything is always streaming. We put Python's async generators center stage and let them shine.
You'll leave knowing how to build a real agent: with tool calls, human in the loop, and durable execution, in plain async Python, and when less framework beats more.
Yury Selivanov is a Python core developer and has been contributing to the language since 2013. He designed and added async/await to Python, maintained asyncio, and created the widely used uvloop and asyncpg libraries. He was co-founder and CEO of Gel (formerly EdgeDB), the open-source database, and is now Director of Engineering at Vercel, where he leads the Python team building tools to make it easier to build and deploy AI agents in Python. He has spoken at PyCon US, EuroPython, PyCon Canada, and many other conferences, and has spent much of his career designing high-performance concurrent systems: the foundation of the AI SDK he's presenting.