中国AI智能体现状分析

The Current State of AI Agents in Industry Update Consumer facing AI agents face significant challengesbecause of lack of access to the walled gardens; enterprises China TechnologyPOSITIVE China TechnologyJiong Shao, CFA+1 212 526 5562jiong.shao@barclays.com The following companies are not under coverage by Barclays Research: [Anthropic, ByteDance,DeepSeek, Kunlun, MiniMax, Moonshot AI, Monica, OpenAI, Zhipu]. Information about thesecompanies is being provided for information purposes only and is not an investment Lian Xiu (Roger) Duan+1 212 526 4633lianxiu.duan@barclays.com Since March 2025, when the first general purpose AI Agent Manus was launched, severalcompanies in China, including tech powerhouses such as ByteDance, Alibaba and Tencent aswell as small startups such as Zhipu, Moonshot AI and Minimax, have launched their AI agents.We tried all these leading AI agents in China and compared them against ChatGPT’s AI Agent. In Xinyao Song+1 212 526 6972xinyao.song@barclays.com With the exception of Manus, all AI agents in China are basedoffeach company's ownproprietary foundational LLMs. On the one hand, we have been arguing that LLMs appear tobe increasingly similar in capabilities and individual LLM may not be the keydifferentiatorforthe popularity or adoption of an AI agent. Having the best LLM may mean very little for targetusers, especially consumers users, who are likely to value other elements such as ease of use far We have seen significant progress made in specialized AI agents in areas such as research (e.g.ChatGPT’s Deep Research) and coding (e.g. GitHub Copilot). However, progress in generalpurpose AI agent has been slow (in the AI sphere). While a specialized AI agent can write code fora complex program in a matter of minutes and ChatGPT’s Deep Research can write a sell-sidetype research report easily, no AI agent today can execute a simple task such as – buy me aMacBook Air 15 on Amazon. This simple task requires no thinking or reasoning, but cannot bedone by the smartest AI agent today for many reasons, the most significant being -why On the other hand, it could be a lot easier for enterprises to adopt AI agents that are of Barclays Capital Inc. and/or one of itsaffiliatesdoes and seeks to do business with companiescovered in its research reports. As a result, investors should be aware that the firm may have aconflict of interest that couldaffectthe objectivity of this report. Investors should consider this workflow, and ERP needs.In the enterprise arena, since the need to access externalproprietary information is quite limited (every AI agent can access external public information),with the help of AI experts (e.g. Alibaba, ByteDance, JD, etc), enterprises can develop a range ofAI agents to improve productivity andefficiencysignificantly. Recent media reports suggest thatalmost 200 enterprises and government entities in China have spent a total of at least RMB1.1bn ($150m) to work with DeepSeek in the first half of 2025 for a variety of projects (discussed Despite significant challenges, especially for the consumer facing AI agents, we believeadoption of AI agents by enterprises and consumers alike may be faster in China than in othermarkets such as the US becausethe inferencing cost in China is roughly 90% below theinferencing cost in the US.As adoption of AI agents takesoff,the demand for inferencing willlikely jump substantially. As in all tech adoption curves, having lower costs tends to turbo CONTENTS Key takeaways from Chinese AI agents for consumers. . . . . . . . . . . . . . . . 4 Stepping back, how does AI Agent work?. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4Key takeaways on Chinese AI agents development. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4AI agent showcase - book a team dinner. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 AI agent development is supported bycost-effectivefeatures, LLM Key takeaways. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12Latest Chinese LLMs innovations and prices. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 At business level, agents are practical due to customization; local governments and SOEs are actively deploying AI models. . . . . . . . . . . . 16 AI glasses - Integration of AI and the physical world. . . . . . . . . . . . . . . . . 21 AI agents and underlying LLMs in detail. . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 Manus (Monica, Not Covered). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21AutoGLM Rumination (Zhipu, Not Covered). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22MiniMax Agent (MiniM