Each small bump in model performance kills another agent framework. The base LLMs can do so much in-context and if they need more internal structure, they can write tools. You don't need the framework. The only thing frameworks should be doing is making the problem easier for the LLM to solve.