Recent discussions highlight concerns that certain agent frameworks are undermining the efficiency of large language models (LLMs) by wasting tokens, which negatively impacts inference speed and cost-effectiveness. Experts suggest that improved collaboration between agent design and inference processes could lead to significant performance enhancements in LLM applications.