I started using agents with AutoGen, then tried Llama-index and Langchain/graph. One thing they all have in common is an opinionated, idiomatic and bureaucratic big idea interface. The whole experience was needlessly frustrating.
Yes, they are trying to encapsulate a common interface over many providers, I understand that, but it does not have to be some big idea based thing.
So a big thanks to the OpenAI folks who worked on the Agents SDK for having done just the right amount of abstractions and shown restraint against the programmer-urge to make something cool, but made a pragmatic SDK that gets out of your way instead.