The rise of agentic AI is shifting the tech landscape from laboratory experiments to systems that can act autonomously across business and personal tasks. At a Shanghai forum on April 3 convened by the Xujiahui Science and Technology Innovation Center, the Shanghai Distributed Consensus Technology Association, PANews and Mankiw Law Firm, speakers outlined both the promise and the practical challenges of that transition.

Li Chenxing, chief architect at Conflux Tree Graph, argued that giving AI greater autonomy is an unavoidable direction for the field, but warned that current systems struggle to retain and apply the contextual constraints required for reliable decisions in complex, real-world environments. He described memory as the chief technical bottleneck, parameter storage, short context windows and slow or inefficient external memory access all limit continuity of experience, and urged work on stronger retrieval, continuous learning and vertical domain practice to build reusable experiential memory.

Practical deployment concerns were a recurring theme. Tencent Cloud’s Feng Heqing explained that mature enterprise agents must support end-to-end tasks, multi-role collaboration and hierarchical memory while preserving data security through local storage and manual confirmation for critical operations. He outlined an enterprise-ready architecture with execution isolation, permission control and both cloud and on-premise deployment options, noting these are necessary to adapt agents to complex corporate workflows.

Speakers with hands-on experience warned that agent systems are still engineering-heavy and resource-sensitive. Teddy, founder of Biteye and XHunt, recommended mandatory multi-stage review processes, such as higher-level agents rechecking code produced by lower-level agents, to reduce errors, and advised careful orchestration of execution via backend APIs to preserve stability and control token consumption. He also highlighted security risks including prompt injection and malicious skill modules.

OpenClaw, the open-source agent framework enjoying rapid uptake in China, was a focal point of discussion. Its plugin-like “skills” enable agents to interact with external services and automate complex workflows, but the community and vendors stress that skills are untrusted code until vetted and that poor or malicious skills can cause substantial harm. China’s fast adoption has been driven by local compute economics and policy incentives, yet regulators have restricted its use in official institutions amid data-security concerns.

The commercial dynamics around OpenClaw have exposed deeper industry tensions. Anthropic recently moved to restrict or monetise third-party agent integrations with its Claude service, citing the disproportionate compute demands of agentic usage versus conventional chat interactions. That policy shift, which includes new usage charges and transitional credits, has provoked pushback from open-source proponents and underscores a broader shift from flat-rate subscriptions toward usage-based pricing for resource-intensive AI applications.

Investors at the event urged sober reading of where durable advantage will arise. Venture capitalists emphasised that rapid model iteration reduces the shelf-life of purely algorithmic leads and recommended concentrating on hard-to-replicate assets such as computing resources, data and user-locked memory systems. Several panellists predicted the emergence of new friction points: whether AI-generated memory becomes portable or product-locked, whether single‑vendor lock-in produces concentrated failure modes, and whether a dominant “super portal” for AI interaction will take hold.

Legal and operational safeguards are already being pressed into service. Mankiw LLP partner Zhao Xuan cautioned entrepreneurs against “false isolation” from one-person corporate shells, urged rigorous documentation to establish ownership of core assets, and recommended designing around platform centralisation risks by separating critical data from third-party services and exploring decentralised options. Such precautions aim to limit single points of failure as agents assume more consequential roles, including transaction execution and strategy implementation.

Source Reference Map

Inspired by headline at: [1]

Sources by paragraph:

Source: Noah Wire Services