🤖 AI Summary
This survey addresses the challenge of systematically understanding and advancing large language model (LLM)-based autonomous agents, particularly their capacity to emulate human learning and decision-making amid persistent interdisciplinary bottlenecks. We propose the first unified taxonomy for LLM-driven agents, rigorously delineating their capability boundaries and establishing principled evaluation paradigms. Methodologically, we conduct an in-depth analysis of the reasoning–action loop, examining core technical challenges—including prompt engineering, tool integration, memory architectures, reflective self-improvement, and reinforcement learning synergies. Synthesizing insights from over 120 representative studies, we identify four dominant research thrusts: architectural design, task planning, environment interaction, and multi-agent collaboration. Finally, we articulate six forward-looking research directions, offering a comprehensive roadmap for both theoretical foundations and practical deployment of autonomous agents.
