Zhou Hongyi, founder of Chinese internet security giant 360 and a member of the national Chinese political advisory body, told reporters that the next step for artificial intelligence is to convert large, general models into autonomous agents that can cooperate across tasks. Zhou argued that familiar large models today function largely as “chatbots” and that only by turning those models into embodied or operational agents — and enabling collaboration among multiple such agents — will AI achieve meaningful industrial adoption.
Zhou framed the debate in grand terms: the original ambition of artificial intelligence was to create an “infinite supply” of human intellect, and its use should therefore not be constrained by individual sectors. His remarks echo a growing consensus in industry and research that the capabilities demonstrated by large language models (LLMs) are an early phase, and that production applications will require systems that can perceive, act, coordinate and persist in real‑world environments.
Technically, the difference Zhou highlights is important. A large language model excels at pattern completion and conversational tasks, but struggles with continuous decision‑making, managing stateful workflows, interacting with external hardware, and handling the complexities of physical environments. Agents — software systems that combine a model with tools, memory, policies and interfaces to sensors and actuators — can decompose complex tasks, call specialised services, and coordinate with other agents to achieve goals that no single model could handle alone.
China’s industrial strategy and private sector ambitions give extra weight to Zhou’s prescription. Domestic cloud providers, robotics startups and enterprise software vendors are racing to translate model research into factory automation, logistics orchestration, customer service automation and healthcare workflows. If multi‑agent architectures prove decisive, companies that can provide orchestration layers, secure data pipes, and trusted integration with industrial control systems stand to capture much of the economic value.
But the path is neither short nor straightforward. Turning models into reliable agents requires progress on perception, long‑horizon planning, safe exploration, simulators for transfer to reality, bandwidth‑efficient orchestration, and governance frameworks for accountability and data protection. Enterprises face additional hurdles: legacy systems integration, regulatory compliance, and the need for measurable return on investment rather than proofs‑of‑concept that look impressive but do not scale.
Zhou’s remarks should be read as both a technical assessment and a strategic nudge. For Chinese industry and policymakers, the message is to move investment and experimentation from isolated LLM research towards engineered systems that combine models, software engineering, robotics and secure cloud infrastructure. Globally, the debate over agents versus models will shape which firms and nations lead the next phase of AI deployment — and how rapidly the technology moves from lab demonstrations into everyday economic use.