the current bottlenecks for agents are actually bottlenecks of the LLMs, rather than the agents themselves.