Nearly 7,000 stores and a large volume of real conversations every day are becoming training material for Infinitus’s AI assistant “Lingzhi.” We have noticed that the key shift is not the deployment of an Agent—an AI program capable of autonomously calling tools to complete tasks—but Alibaba Cloud AgentLoop’s ability to process operational logs into an evaluation-and-feedback loop for continuous improvement.
What this is
Infinitus relies on three dedicated knowledge bases to support store partners with product materials, sales scripts, team coaching and service handling. Previously, the system recorded the full chain of every response, but mainly treated it as logs: problematic samples depended on manual searches, while lessons learned from modifying Prompts—instructions that define a large language model’s task—remained in the hands of individual employees.
The new approach breaks the chain into seven stages: observability integration, data processing, evaluation, human annotation, data feedback, experience mining and experience integration. User feedback is first linked to a specific answer and its execution chain, after which the system scores it for accuracy, relevance, safety and tool-call correctness. Low-scoring samples enter human review, and approved experience is then fed back into the Agent.
Alibaba Cloud’s AgentLoop handles data processing and experience production, while Infinitus’s in-house “Lingshu” manages human annotation and review. This division of labor means the system is not fully autonomous self-evolution; it embeds human judgment into a traceable process.
Industry view
This releases a clear signal: the focus of corporate competition in AI applications is shifting from “whether it can answer” to “whether it can continuously improve.” For companies with large volumes of real business conversations, online trajectories are themselves scarce quality data. Linking complaints, incorrect tool calls and excellent answers to their underlying processes is closer to actual business value than simply increasing model parameters.
But the loop is not inherently reliable. The cost of Alibaba Cloud’s solution also comes primarily from data governance: a Trace—a structured record of one Agent execution—must undergo extraction, anonymization, deduplication and standardization, while mapping feedback to trace fields involves preserving context across root and child nodes. If automated evaluation is biased, it may write incorrect experience back into the system; if human annotation lacks unified standards, scaling it can create a new bottleneck.
Therefore, “Infinitus” in the solution’s name carries a degree of product marketing. Whether AgentLoop can be replicated across other enterprises depends on the quality of business data, knowledge-base governance, evaluator accuracy and the organization’s execution capabilities—not merely platform integration.
Impact on regular people
For enterprise IT: When enterprises procure Agents, evaluation will increasingly shift from demo performance to observability, data feedback and governance. Solutions that can explain how failed samples are handled and how learned experience is reviewed will be better suited to core business operations.
For individual careers: Store sales staff, customer service representatives and operations employees will gradually work with more proactive AI assistants. In the short term, the key issue is not “being replaced by AI,” but that work results, effective sales scripts and customer feedback will be easier for systems to capture, reuse and scale—potentially widening the gap between high- and low-performing employees.
For the consumer market: Every time a consumer gives an AI assistant a negative experience, that feedback may trigger a complete quality-tracking and improvement process. As similar mechanisms enter e-commerce, customer service and local services, AI will move beyond generic answers toward continuous optimization tailored to individual enterprises.