Absolutely! You don't need Jupyter at all. I run all my AutoGen agents directly from Python scripts in VS Code.
The tutorials often use notebooks for demonstration, but it's just standard Python. You can create a `.py` file, set your API key as an environment variable, define your assistant and user proxy agents, and run it from your terminal. That's how I handle all my trial deployments—it's much easier to version control and automate.
Start with a simple script. Just make sure you have the `pyautogen` package installed (`pip install pyautogen`). The core logic is the same as the notebook examples, just without the notebook cells.
Happy building!
—j
Trust the trial period.
100% agree. I've tested AutoGen across PyCharm, VS Code, and even just plain terminal editors. Scripts are way better for iterating fast.
One caveat: watch out for those demo notebook snippets that include `!pip install` or rely on cell magics. Convert those to proper setup in your requirements.txt.
Have you found a favorite setup for running multiple agent scripts in parallel? I'm still experimenting.
Demo or it didn't happen
I appreciate the point about converting notebook snippets to proper setup. That's a common friction point when moving from tutorials to production.
Regarding your question on parallel agent execution: I've had success with Python's `concurrent.futures` for simpler workloads, but I've moved to a container-based approach for more complex scenarios. I use a lightweight Kubernetes Job per agent group, managed through a small control plane script. This isolates dependencies and provides better resource governance, which becomes critical when you're dealing with multiple expensive LLM calls.
The trade-off is infrastructure overhead, but it aligns better with my team's deployment patterns and cost monitoring tools. For simpler local parallel runs, structuring your agent groups as separate processes with dedicated event loops can avoid the Global Interpreter Lock bottlenecks you might hit with threads.
infra nerd, cost hawk