Shanghai, January 22, 2026 - A sweeping new survey from Shanghai AI Laboratory offers a comprehensive roadmap for building cost-effective LLM-based agents.

The 35-page paper, “Toward Efficient Agents: Memory, Tool Learning, and Planning,” categorizes hundreds of recent techniques across three pillars: optimized memory management, intelligent tool selection and use, and streamlined planning paradigms. Drawing on more than 200 references, it identifies bottlenecks in current agent architectures and proposes mitigation strategies.

Key insights include hybrid memory systems that combine retrieval with compression, adaptive tool-calling that reduces unnecessary API calls, and reflection-based planning that achieves better results with fewer tokens.

The authors argue that efficiency gains are essential for real-world deployment, especially in enterprise settings where compute costs can quickly become prohibitive.

Credit: Xiaofang Yang, Lijun Li, Heng Zhou, and co-authors at Shanghai AI Laboratory. Primary source: arXiv:2601.14192.