Wanlian Zhice has configured roughly 800 agents in production and used 39 days to evolve “Xiao Tu” from a conversational assistant into a team of digital workers. Our judgment is that the competitive focus of enterprise AI is shifting from “whether a model can answer” to “whether agents can continuously complete business tasks.” But scale figures still need to be assessed against actual usage rates.
What This Is
Wanlian Zhice is Wanlian Yida’s enterprise operations decision-making application, covering opportunity discovery, supplier research, risk analysis, and industrial-chain research. In this partnership with Alibaba Cloud, it uses AgentCore as the runtime foundation for its multi-agent system and AgentLoop—a tool for observing, evaluating, and optimizing agent task execution—to track execution quality.
The system follows a Leader–Worker model: one Leader asks follow-up questions, breaks down tasks, and consolidates results, while four categories of specialist agents handle sales, supply chain, risk control, and strategy. They connect to Wanlian Zhice’s existing data and applications through skills and MCP, the interface protocol that lets agents invoke external tools and business systems. Critical actions such as purchase orders and contract signing still require human confirmation.
One distinction is essential: the roughly 800 agents represent the scale of business configurations in production, not 800 compute instances running simultaneously. The system starts and stops agents by task, reducing idle costs, but this also means the deployment count cannot be equated directly with usage scale or business impact.
Industry View
What stands out is that this implementation no longer centers on single-turn question-and-answer exchanges. Instead, it divides work into complete tasks that can be assigned, tracked, and evaluated. Backtesting with fixed question sets, online experiments, and task-level cost accounting are also much closer to the operating mechanisms enterprises actually need than simply comparing final answers.
The risks are just as clear. The more steps a multi-agent system contains, the more errors can accumulate across task allocation, tool calls, and result synthesis. The accuracy of business data, isolation of permissions, and the ability to review critical conclusions all affect final reliability. Moreover, completing a production launch in 39 days demonstrates engineering speed, not long-term task success rates, cost savings, or return on investment. Alibaba Cloud is both the infrastructure platform provider and an important driver of this case; the data still has to withstand scrutiny based on actual business results.
Impact on Regular People
For enterprise IT: As organizations procure AI services, they will increasingly prioritize integration with existing systems, permission controls, process auditing, and task-level billing—not just model parameters. The core value of platforms such as AgentCore is to connect agents directly to business operations.
For workplace professionals: Sales, procurement, and risk-control staff are more likely to work in a model where “AI prepares the first draft and research, while people verify and decide.” In the near term, process transparency and traceable results matter more than the number of digital workers.
For consumers: Consumers will not immediately see 800 agents, but they may encounter the effects through supplier recommendations, risk alerts, and faster service responses. Without human review, agents may also propagate errors into purchasing decisions more quickly.