Today, the agent economy is no longer science fiction; it brings not only a leap in efficiency but also a restructuring and redistribution of economic organization. In particular, the global popularity of the open-source project OpenClaw has further propelled large-scale models from the laboratory to large-scale applications, leading to a fierce competition among various parties to enter the agent market.
So, which platform should we choose for the large-scale model? Are the token resources sufficient to support long-term use? Will we be left behind if we don't follow the OpenClaw trend? In this rapidly evolving AI revolution, how should individuals position themselves and break through?
With these questions in mind, on April 3, Xujiahui Science and Technology Innovation Center, Shanghai Distributed Consensus Technology Association, PANews, and Mankiw Law Firm jointly held an event themed "Don't Be Anxious About Shrimp".
In his keynote speech on "Embracing the Unpredictable AI Wave," Li Chenxing, Chief Architect of Conflux Tree Graph, stated that granting AI more autonomy, rather than excessively constraining it with limited human experience, is an inevitable trend at this stage of technology. The "lack of consideration" currently exhibited by AI essentially stems from its difficulty in consistently capturing and continuously remembering key contextual constraints in complex scenarios. From a technical perspective, AI primarily relies on parameter memory, contextual memory, and external memory, but these mechanisms still suffer from difficulties in updating, limited windows, and insufficient invocation efficiency. Therefore, future efforts should focus on strengthening external memory retrieval capabilities, exploring mechanisms for continuous learning and experience reuse, and gradually accumulating experiential memory through vertical domain practices to improve the completeness and reliability of AI's decision-making in real-world complex scenarios.
He also pointed out that the core progress of AI currently lies mainly in the enhancement of its autonomous analysis and reflection capabilities. In the future, with improvements in memory capacity, it is expected to break through key bottlenecks and have a profound impact on various industries. For example, the potential of digital identity and digital payment systems has long been constrained by development and user barriers, while AI is expected to unlock its value by reducing development costs and replacing the user learning process with an agent-based approach. Overall, AI should not be seen as a threat to employment, but rather as a key tool for driving productivity improvements and creating new opportunities. Individuals and industries should maintain an open mind and proactively explore paths for AI integration.
According to Feng Heqing, product architect of Tencent Cloud Workbuddy, with the significant improvement in large-scale model capabilities, AI has evolved from supporting only basic development assistance such as code completion to being able to independently complete complex tasks. The core capabilities of custom agents are reflected in full-process task support, multi-role collaboration, a hierarchical memory system, and context-based intelligent task decomposition. Simultaneously, multi-agent collaboration enables data flow and parallel processing between tasks, and local data storage and manual confirmation mechanisms for critical operations ensure data security. At the application level, WorkBuddy covers typical office scenarios such as resume screening, automatic PPT generation, data analysis, and weekly report integration. It can also connect to systems like WeChat Work through enterprise-level integration capabilities for unified task management. Its technical architecture emphasizes full-stack self-development, execution environment isolation, and enterprise-level permission control, supporting both local and cloud deployments. In terms of business model, it can target enterprise R&D and users in high-frequency digital office positions. Overall, WorkBuddy aims to improve enterprise productivity through custom agents and multi-task collaboration capabilities, and further strengthens its adaptability and implementation capabilities in complex enterprise scenarios by continuously optimizing task decomposition capabilities and expanding its ecosystem.
Teddy, founder of Biteye and XHunt, primarily shared his insights on digital employee practices, large-scale model applications and cost issues, technology configuration and security risks, and optimization of collaboration methods. Regarding digital employee practices, to reduce model illusions and code error rates, a higher-level review agent is needed to conduct a secondary review of the code generated by lower-level agents, establishing a mandatory code review process. Since current agent-based code still has some bugs, errors can be reduced by standardizing development processes, strengthening prompt word design, and adding multi-round verification mechanisms. In operational scenarios, it's crucial to control posting frequency and ensure stability through unified scheduling via backend APIs. In complex team collaboration environments, Discord is generally more suitable than Telegram for agent collaboration and task distribution, and special attention needs to be paid to token consumption in resource management. Furthermore, agent systems still require human investment for training, optimization, and behavior correction.
Regarding the installation and deployment of OpenClaw, Teddy suggests running it on an idle computer or Mac Mini, offering a high degree of autonomy, open-source code, strong privacy protection, and integration with an international ecosystem. However, its installation and configuration are relatively complex. Special attention should be paid to the risks of modifying model and channel configurations to avoid system anomalies due to improper configuration. Tools such as Grok and Gemini can be used to assist in troubleshooting when problems arise. Security-wise, it's crucial to guard against risks such as prompt word attacks and malicious skill injection. In terms of resources and costs, token consumption control is also necessary to avoid excessive operating costs.
In his keynote speech, Zhao Xuan, a partner at Mankiw LLP, shared three major legal issues and solutions that entrepreneurs need to pay attention to in the AI era. The first is the organizational shell, namely the "false isolation" created by one-person companies (OPCs). While they appear to be independent entities, they are ineffective in truly isolating responsibility and risk. Genuine physical and legal separation is needed, including introducing partners in the structure, using dedicated corporate credit cards, and including AI disclaimers and compensation caps in contracts. The second is the issue of core asset ownership. Effort does not equal rights; it is necessary to prove one's control, fully record and document the creative process. The third is the systemic risk of "pulling the plug" brought about by platform hegemony, including "god-like" clauses and technology lock-in. This requires separating core data from third-party services, planning alternative solutions in advance, and introducing decentralized technology.
In the roundtable discussion titled "From Frenzy to Sobering Up: The Real Needs and False Propositions of AI in the Eyes of VCs," several investors shared their insights on the development stages, application boundaries, and investment logic of AI.
Cancer, founding partner of Waterdrop Capital, believes that AI is still in its early stages of development, and it will take a considerable amount of time to reach a stage where user experience is mature and it is widely considered "meaningful." He points out that AI technology iterates extremely rapidly, and simply relying on technological leadership is insufficient to create a long-term competitive advantage. Therefore, investments should focus more on fundamental capabilities with irreplaceable value, such as core resources like computing power. At the application level, he cites the example of tools like "Lobster," which are not user-friendly for ordinary programmers, but may be better suited for future development into vertical applications such as "family doctor," providing professional advice through real-time health data. He also believes that AI can replace information production tools like research reports in enterprises, but it cannot replace the final decision-making role; it can only exist as an auxiliary decision-making tool.
Tang Yi, founding partner of Enlight Capital, stated that it's currently difficult to identify clear non-consensus opportunities in the AI investment field, as the rapid iteration of large-scale models may continue to "level out" the advantages of application-layer companies. He is relatively optimistic about the combination of Web3 and AI, believing that they represent advanced productivity in their respective fields. Regarding open-source tools like OpenClaw, he believes they essentially give large-scale models "hands" and "feet," enhancing their connectivity with external systems and social applications, but also bringing higher security and data risks. Therefore, they require complex configurations and are not suitable for ordinary users. Currently, a more ideal approach is to improve overall usability and experience through encapsulation.
First Rule Ventures investor Yinghao focuses on application opportunities in deep-water industry applications, AI creation, and the integration of software and hardware from a user and product perspective, and evaluates project potential through user behavior and interaction data. He points out that even if you don't personally try all emerging AI products, it doesn't mean you'll miss key trends, because technological capabilities are often quickly modularized and integrated into existing product systems.
Compared to a single product, he is more concerned about three long-term structural changes: First, whether AI interaction is forming a new memory carrier, allowing users' cognition and work to be deposited in a certain system; second, whether this memory has the ability to migrate across products, or will gradually be bound to a single product, thus forming high migration costs and experience lock-in; and third, whether a new super portal will emerge, becoming the core hub for AI interaction and traffic distribution.
Zhao Xuan, a partner at Mankiw LLP, primarily uses AI products for data processing, retrieval, and analysis, and anticipates the emergence of more integrated products that combine these capabilities. He also emphasizes that avoiding a single major failure is crucial in AI startups, advising companies to prioritize key legal design elements such as data compliance, arbitration clauses, and disclaimers from the outset. This aims to isolate risks and protect liability in the event of uncontrollable risks, preventing a single point of failure from causing the company's collapse. Furthermore, he envisions agents becoming the primary economic actors in the future, responsible for data acquisition, information purchase, strategy execution, and even cross-system transactions, thus forming a machine-to-machine economic activity and payment system.
In a roundtable discussion themed "N Ways to Unlock AI: Opportunities for Innovators," several guests explored the changes brought about by AI from different perspectives. Zeno, CEO of Matrix Intelligence, proposed that users can connect multiple devices by modifying scripts or plugins, achieving multi-device memory synchronization and state consistency, ensuring no information loss and uninterrupted task flow. Daily cleanup/review mechanisms can also be added to maintain system stability. Compared to using off-the-shelf tools, deep customization based on enterprise-level permissions or platform capabilities is more efficient, more flexible, and easier to create workflows tailored to individual habits. Looking to the future, he believes AI will become a unified entry point, allowing users to interact through a single AI hub to access various tools and systems to complete all tasks. As usage increases, AI will continuously accumulate user memories, preferences, and workflows, creating a flywheel effect of data and capabilities, becoming increasingly user-savvy and efficient. Under this trend, individuals, by configuring AI systems and paying subscription costs, may achieve productivity gains far exceeding traditional human labor, significantly widening the efficiency gap between individuals.
ClawFirm.dev co-founder 0xOlivia revealed that in practical AI applications, issues such as system instability and fragmented memory and automation capabilities still exist. Users need to continuously assemble various tools and scripts, much like building with Lego bricks. For non-advanced users, directly adopting mature commercial platforms combined with official applications and continuous iteration capabilities is often more stable and efficient than highly fragmented self-built systems. Furthermore, introducing open-source components can further enhance data processing and content generation capabilities. She emphasized that the main limitation of AI currently lies not in the model's capabilities themselves, but in the fact that its engineering application methods have not yet fully matched those capabilities, thus leaving enormous room for optimization and implementation. In the future, as the capabilities of large models rapidly increase, AI applications will gradually cover all aspects of work and life, and continue to integrate with different product forms.
When discussing AI digital employees, Biteye/XHunt founder Teddy pointed out that AI can be integrated into internal systems via APIs or automation interfaces, enabling it to perform specific execution tasks such as code generation, requirement fulfillment, and content processing. Humans, on the other hand, can focus on product design and requirement definition, thus retaining key decision-making power. This collaborative model is more stable and scalable, not only improving overall development efficiency but also significantly reducing error rates, making AI more like a schedulable and manageable outsourced team rather than a single tool. He also emphasized that any process-oriented and repetitive task has the potential to be transformed or replaced by AI. Even if the initial effects are unstable, long-term optimization and gradual enhancement of productivity are possible. In complex tasks and management decision-making, AI has also begun to demonstrate significant auxiliary capabilities and is penetrating into higher-level business scenarios.
Dou Ge, a senior AI application development engineer, added that while there is a general consensus on the trend of AI outsourcing, automation, and tool-based collaboration, from a business perspective, it is even more important to consider security, access control, employee collaboration mechanisms, and asset accumulation. Currently, the market offers various AI development frameworks and tool ecosystems, each with its own focus on lightweight design, low-code implementation, high integration, and security controls. Enterprises need to strike a balance between flexibility and controllability when choosing these frameworks and design their architectures in accordance with their actual business scenarios. Truly understanding and implementing these AI systems cannot remain merely theoretical; it requires actual investment and usage costs. He emphasized that AI is rapidly reshaping workflows and organizational structures. Both individuals and businesses must quickly adapt to this change, improving efficiency through continuous learning and tool-based applications; otherwise, they risk being left behind by the rapid pace of technological iteration.
