The rapid progress of AI agents has been driven by breakthroughs in foundation models, training algorithms and system design. In this talk, I will introduce AGILE, a general reinforcement learning framework for large language model agents. Building upon AGILE, we have developed a series of deployable agents: PaSa, an academic paper search agent that significantly outperforms mainstream search engines such as Google and Google Scholar; and M3-Agent, a multimodal agent with long-term memory that can continuously process real-time multimodal inputs, incrementally build world knowledge, and reason over its memories. Finally, I will share our latest research on the mechanisms of AI agents, focusing on memory retrieval and consolidation in large language models. Together, these explorations aim to shed light on how future AI agents can become more autonomous, intelligent, and trustworthy.