The rapid evolution of Agentic AI marks a new phase of artificial intelligence, where Large Language Models (LLMs) no longer merely respond, but act and adapt. This talk traces a paradigm shift from workflow-based agent systems, in which planning, tool use, and memory are coordinated by external logic, to the emerging model-native paradigm, where these capabilities are increasingly internalized within the model. Reinforcement Learning (RL) is presented as the key driver of this transition, enabling outcome-driven learning beyond static imitation and supporting a unified “LLM + RL + Task” perspective across language, vision, and embodied settings. Under this, planning, tool use, and memory are reviewed as they evolve from externally scripted modules into end-to-end learned behaviors. The talk further examines how this shift reshapes major applications such as Deep Research agents for long-horizon reasoning and GUI agents for embodied interaction. Overall, these trends outline a coherent trajectory from engineering workflows that apply intelligence to building models that develop intelligence through experience.

Jitao Sang is a Professor at Beijing Jiaotong University. His research interests include multimedia content analysis, trustworthy and AI alignment, AI Agent. He has received several awards including the CAS President’s Special Award and the ACM China Rising Star. He has published over 50 CCF-A papers, with first- and second-author papers have received 8 paper awards in CCF-recommended conferences.