Alibaba’s September 22 full-stack agent roadmap, including the AgentCore platform, agentic cloud infrastructure, Qwen Intelligence mobile agents, and claims of recursive self-improvement on Qwen3.8 Max, has driven its market-implied lead for third place among AI agent labs by year-end. Strong open-weight benchmark results on agentic tasks such as Terminal-Bench and OSWorld position Alibaba competitively against higher-cost U.S. models, while Tencent’s security-focused agent research and other Chinese labs like Moonshot and MiniMax add depth to the regional push. U.S. labs including OpenAI (Dots always-on agents at DevDay), Anthropic (Claude 5.5 series with strong coding and long-horizon performance), and Google (Gemini voice agents and Skills) maintain frontier edges but face higher pricing and slower open-weight adoption. With probabilities clustered tightly, rapid model iterations and benchmark shifts through December remain key swing factors.