返回首页

对比阅读

对比阅读:Alibaba Cloud AgentLoop Turns Store AI Logs Into Experience, but Humans Still Needed 与 阿里云 AgentLoop 让门店 AI 从日志走向经验,但闭环仍需人工把关

AEN
Alibaba CloudAgentLoopInfinitus·

Alibaba Cloud AgentLoop Turns Store AI Logs Into Experience, but Humans Still Needed

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.

来源: juejin.cn
BZH
阿里云AgentLoop无限极·

阿里云 AgentLoop 让门店 AI 从日志走向经验,但闭环仍需人工把关

近 7000 家门店、每天大量真实对话,正在成为无限极 AI 助手“灵知”改进能力的训练素材。我们注意到,关键变化不是 Agent(能自主调用工具完成任务的 AI 程序)上线,而是阿里云 AgentLoop 把运行日志加工成可评估、可回流的持续改进闭环。

这是什么

无限极依托三类专属知识库,为门店伙伴提供产品素材、销售话术、团队辅导和业务办理支持。过去,系统记录了每次应答的完整链路,但主要被当作日志使用;问题样本依赖人工翻查,Prompt(给大模型设定的任务指令)修改经验也留在个人手里。

新的做法是将链路拆成七个环节:可观测接入、数据处理、评估、人工标注、数据回流、经验挖掘和经验接入。用户反馈先关联到具体回答及其运行链路,再由系统按准确性、相关性、安全性和工具调用正确性打分。低分样本进入人工审核,通过后的经验再回接 Agent。

阿里云的 AgentLoop 承担数据加工和经验生产,无限极自建的“灵枢”负责人工标注与审核。两者分工意味着,这不是完全自动的自我进化,而是把人的判断嵌入一条可追踪的流程。

行业怎么看

这件事释放了一个明确信号:企业竞争 AI 应用的焦点,正从“能否回答”转向“能否持续改进”。对于拥有大量真实业务对话的公司,线上轨迹本身就是稀缺的质量数据;能否把投诉、错误调用和优秀答案对应起来,比简单扩大模型参数更接近实际业务价值。

但闭环并不天然可靠。阿里云方案的成本也主要来自数据治理:Trace(一次 Agent 执行过程的结构化记录)必须经过抽取、脱敏、去重和标准化,反馈与链路的字段映射还涉及根节点和子节点的上下文传递。自动评估若存在偏差,可能把错误经验重新写回系统;人工标注若缺少统一标准,规模化后仍会形成新的瓶颈。

因此,“无限极”在方案名称中带有产品宣传色彩。AgentLoop 能否复制到其他企业,还取决于业务数据质量、知识库治理、评估器准确率和团队执行能力,而不是只靠平台接入。

对普通人的影响

对企业 IT:企业采购 Agent 时,评估重点会从演示效果转向可观测性、数据回流和治理能力。能够说明失败样本如何被处理、经验如何审核的方案,更适合进入核心业务。

对个人职场:门店销售、客服和运营人员会逐渐面对更主动的 AI 助手。短期内,关键不是“被 AI 替代”,而是工作结果、优质话术和客户反馈更容易被系统沉淀与复用,经验型岗位的差距可能因此拉大。

对消费市场:消费者向 AI 助手反馈一次差体验,背后可能对应完整质量追踪和改进流程。类似机制进入电商、客服与本地生活后,AI 服务会从通用回答转向更贴近具体企业的持续优化。

来源: juejin.cn