Agents are still too advanced. I feel that before building a true agent, the model’s memory/architecture should be redesigned. Agents are at most an interim solution; the ultimate direction of LLMs should still be integrating other apps.