At the 13th Internet Security Conference (ISC.AI2025) held on August 5, Zhou Hongyi, founder of 360 Group, shared important insights on the development of artificial intelligence. He drew inspiration from the classification of autonomous driving intelligence levels and proposed a complete L1 to L5 classification model for intelligent agents for the first time, clarifying 360 Group's strategic positioning within it.

Zhou Hongyi pointed out that the highest level, L5-level intelligent agent, will have the ability to "create intelligent agents" and represents the ultimate form of AI development.

QQ20250806-114822.png

He divided current and future intelligent agents into the following levels:

  • L4-level - Multi-agent Swarm: Capable of collaborating to complete complex tasks, such as generating long videos or developing websites, which are creative works. The tokens consumed per task are large, ranging between 10 million to 100 million.

  • L3-level - Reasoning Intelligent Agent: Represented by models like Manus, they possess deep research and analytical capabilities, capable of handling complex tasks such as creating professional posters and web pages. The tokens consumed per conversation are approximately 1 million.

  • L2-level - Workflow Intelligent Agent: Primarily applied in scenarios such as knowledge base management and process automation, with token consumption per conversation less than or equal to 1 million.

  • L1-level - Chat Assistant: Represented by OpenAI's ChatGPT, it is mainly used in areas such as emotional companionship and basic Q&A.

In his sharing, Zhou Hongyi emphasized that 360 Group's goals go far beyond the L2 level. He revealed that the company is working on developing high-level intelligent agents at the L3 and L4 levels. To this end, 360 has launched the "Intelligent Agent Factory," a platform that supports the construction of intelligent agents and intelligent agent swarms from L1 to L4 levels. The aim is to allow every enterprise and individual to customize their own L3-level reasoning intelligent agents and L4-level intelligent agent swarms according to their specific needs, thereby gaining stronger productivity in the AI era.