From the release of the general AI (artificial intelligence) agent Manus, which ignited the market, to the continued popularity of the open source AI agent OpenClaw, commonly known as "lobster", to the intensive deployment of desktop agent tools by Internet enterprises such as Alibaba and Baidu, AI agents are moving from the code world to work and life scenes at an unprecedented speed. While the visions of "7 × 24-hour digital employees" and "one person company" are accelerating into reality, recent policies have been introduced in Shenzhen, Wuxi and other places to support the implementation and application of open-source intelligent agent projects such as OpenClaw in manufacturing, government and other fields. However, challenges facing technological maturity, security governance, and industrial ecology have also emerged. Industry experts believe that AI intelligent agents are in a critical "leapfrog period" from technological exploration to large-scale applications. At present, it is more like a 'junior intern' who needs careful debugging. Unlike conversational AI such as ChatGPT and Doubao that users are familiar with, recently popular executable agents such as OpenClaw have essential differences in their technical positioning. AI agents can be understood as digital employees who can both think and execute. ”Zheng Jichang, Dean of the China Innovation Service Research Institute at Tsinghua University, stated in an interview with the Economic Reference Daily that the core difference between the two lies in whether they possess "closed-loop execution capability". AI agents can integrate with various large models such as ChatGPT for deep thinking and content generation. Unlike traditional AI assistants, they also have "hands-on ability". Zheng Jichang pointed out that AI agents can directly convert instructions into actual operations. By invoking system permissions and automation tools, it can perform specific operations on the computer like a human. Zheng Jichang gave an example that when a user requests the task of "organizing this week's work and generating a weekly report, and then sending it to the department's Feishu group", the AI agent can automatically read the relevant files to generate documents, and open Feishu to complete the sending, achieving a complete process from understanding the task to completing execution. However, from a technological perspective, the statement that 'AI starts acting on its own' still needs to be viewed more rationally. Zheng Jichang pointed out that currently, AI agents are more like "junior interns" who need to be carefully debugged. Although they can efficiently complete tasks in specific scenarios, they often have misunderstandings or operational errors. There is still a considerable distance to go before the public expects stable and reliable autonomous intelligence. Since the launch of the commercial exploration AI agent, cloud enabling has accelerated the layout of the industry in many ways. Domestic "cloud" manufacturers, including Tencent Cloud, Alibaba Cloud and Baidu AI Cloud, have access to it. By simplifying the deployment process, providing a cloud operating environment and other measures, it helps users lower the threshold of use. Institutions predict that the size of China's AI intelligent agent market will exceed 3.3 trillion yuan by 2028, and the penetration rate of enterprise level applications will rapidly increase. Taking the open-source desktop agent tool CoPaw launched by Alibaba on February 28th as an example, users can not only deploy locally and in the cloud with just one click, but also conduct secondary development based on CoPaw, freely integrate local models, write skills, and access exclusive messaging applications to meet more customized scenario requirements. According to the CoPaw team, the tool supports access to various communication platforms such as DingTalk, Feishu, QQ, Discord, and has built-in modules for document editing, news reading, and file management. The system can also use task scheduling mechanisms to enable intelligent agents to automatically organize emails, generate weekly reports, or manage to-do items. The Copaw team stated that "in the future, we will further explore the mechanism of collaboration between large and small models, allowing lighter local models to handle private data, and more powerful cloud models to handle planning and coding, while balancing security and performance." Industry insiders believe that the emergence of such tools is lowering the threshold for the use of AI agents, providing a foundation for more developers and enterprises to explore application scenarios. In the view of Zhu Keli, the founding director of the National Research Institute for New Economy, AI intelligent agents are currently in a stage of clear concept, clear direction, concentrated hotspots, and huge potential, but their maturity still needs to be improved. At the technical level, planning ability, tool calling ability, and multi-step task execution ability are rapidly improving, but there is still significant room for improvement in the stability of complex tasks, the reliability of long chain logic, and the ability to cross scenario universality. At the industrial level, large-scale commercial use is still mainly concentrated in automation tasks with relatively simple structures and low risks. For the next development of AI intelligent agents, Zheng Jichang believes that AI intelligent agents should follow a robust path from single point application to system collaboration, from efficiency improvement to decision penetration. By 2026, AI intelligent agents will enter a period of large-scale implementation and governance exploration. At the same time, he emphasized that companies need to quickly realize the need to prioritize building an "AI ready" data base and governance framework. From 'usable' to 'reliable', large-scale applications still need to break through. The popularity of OpenClaw is not only a natural result of technological evolution, but also a stress test of the current state of the industry. How to bridge the gap from "usable" to "reliable" has become the key to determining whether AI agents can truly move towards large-scale applications. As AI agents begin to integrate into critical business environments such as email systems, code repositories, and databases, their potential risks are gradually becoming apparent. 360 Digital Security Group security experts have pointed out that due to improper configuration of OpenClaw agents during installation, deployment, and use, some typical security risks have already emerged. For example, attackers may embed hidden instructions in webpage content through "prompt word injection". Once the intelligent agent reads the relevant webpage, it may be induced to leak sensitive information such as system keys; In practical applications, if the intelligent agent has a deviation in understanding user instructions, it may also pose a risk of misoperation, such as accidentally deleting emails or core business data. Zheng Jichang believes that intelligent agents face a key bottleneck in moving from "usable" to "reliable", and breaking through this bottleneck requires efforts from three directions. Firstly, there is infrastructure reform. He stated that it is necessary to build a high-performance and highly secure "operating base" for intelligent agents. Secondly, architecture evolution is to promote the closed-loop evolution of agents from a single model to "perception reasoning execution self evolution". Finally, regarding governance and protocols, he suggested establishing "security rules" for intelligent agents and establishing a unified security and accountability framework, such as requiring agents to inherit user permission boundaries (identity propagation) in any operation, to achieve full chain traceability. Only by synchronously breaking through infrastructure, architecture design, and governance rules can intelligent agents truly win user trust and become reliable executors in the digital world. ”Zheng Jichang said. Zhu Keli also emphasized that efforts should be made from multiple aspects such as technical architecture, task design, and verification mechanisms. Strengthen planning and reflection abilities, enabling intelligent agents to have self checking, error correction, and dynamic adjustment capabilities during execution; Establish a human-machine collaboration mechanism and retain manual confirmation and supervision at key operational nodes; Improve industry standards and testing systems to form quantifiable and verifiable reliability indicators. (New Society)