Rethinking of the code generation approach

[hughlv]

Even though code generation is the looked best way to keep the tool useful and independant to runtimes other than AutoGen, I've seen the problems such as the efficiency and observability issues. Therefore, I'm considering to make it a workflow engine instead of a standalone program. This will introduce significant changes to backend project.

Any thoughts pls let me know.

[shanmu-raoyunfei]

langflow's integration of crewAI is worth learning from.

[hughlv]

langflow's integration of crewAI is worth learning from.

Thanks for the recommendation. I'll look into it when I get time.

[gofullthrottle]

From my general perspective, I have envisioned an agentic system in which many tasks & goals all go through a somewhat similar process. They are all varying configurations of the same underlying lego pieces. So Code Generation is just one of many agentic teams, and other workflows are executed via their own configurable team(s).

I can see your point in terms of concerns regarding efficiency and observability. Would you mind elaborating a bit on what differentiating factors (with regards to implementation) you see which make way for this fork in the road?

great work btw with the interfaces etc, you plucked some visions right out of my mind ;-)