China pushes AI agent governance and multi‑modal model optimization

China’s Ministry of Industry and Information Technology, together with the Cyberspace Administration and the National Development and Reform Commission, issued the "Implementation Opinion on the Standardized Application and Innovative Development of Intelligent Agents". The document defines an intelligent agent as a system with autonomous perception, memory, decision‑making, interaction and execution capabilities, and places agent governance on the national policy agenda. It distinguishes the term “autonomous” from the earlier ISO/IEC notion of an “automated entity” and calls for coordinated oversight involving state bodies, enterprises and research communities. The opinion arrives amid global concerns about advanced AI models, highlighted by Anthropic’s Claude Mythos and reports from the UK ICO and the Five‑Eyes intelligence alliance on safety and financial‑system risks.

At the AICon Shenzhen 2026 conference, industry leaders from Tencent, Alibaba, Huawei, Baidu and Ant Group discussed how to move AI from isolated model performance to scalable intelligent systems. A key presentation by Inferact researcher Mo Zifeng detailed vLLM’s recent optimizations for multi‑modal large language models, including encoder cache, encoder data‑parallelism and CUDA‑graph techniques that reduce latency and cost. vLLM now supports major multi‑modal models such as Kimi‑K2.6 and Qwen‑3.6 and is deployed by cloud providers including AWS, Meta and Digital Ocean. The conference emphasized AI infrastructure, system‑level engineering and agent safety as critical future directions.