当人工智能真正进入企业生产环境，衡量技术价值的标准已不只是模型能力，更在于能否稳定运行、精细调度、安全运维并持续进化。近日，第五届“鼎新杯”数字化转型应用征集成果公布，阿里云云原生应用平台团队四项实践案例获奖，覆盖智能体运行、大模型流量治理、数智运维和智能体全生命周期优化，集中展现支撑推动企业级AI应用规模化落地的工程能力。
第五届“鼎新杯”由中国通信标准化协会主办、中国信息通信研究院承办，自2026年05月正式启动，历时近 5 个月，聚焦国内政企数智化转型领域的创新成果，旨在为行业提供可复制、可推广的实践参考。
四项创新实践获奖案例，覆盖智能体生命周期
Qoder Cloud Agents 基于 RocketMQ LiteTopic 的大规模智能体托管与运行实践
让智能体大规模“跑起来”，首先要解决长程任务、高并发会话与算力成本之间的矛盾。阿里云 Qoder Cloud Agents 基于 RocketMQ LiteTopic 构建“手脑分离、事件驱动”的云端运行架构：智能体负责决策，执行节点按需领取任务、完成后即释放资源；每个会话拥有独立的有序执行通道，故障后可由其他节点无损接续。目前，该方案已支撑万级会话并发，实现“会话常驻、计算流动、状态可续”，将消息中间件从异步通信工具拓展为智能体运行底座。
RocketMQ LiteTopic 支撑大模型服务平台精细化流量治理
面对百万级租户同时调用大模型，传统限流只能决定“放行还是拒绝”，却难以兼顾稀缺算力的平稳利用与租户体验。阿里云大模型服务平台百炼基于 RocketMQ LiteTopic，以“用户＋模型”为维度构建分布式漏桶，将限流重构为精细化节奏调度。海量租户共享资源池，单个租户的拥塞不会影响其他租户。生产实践中，单个消息代理节点可承载 200 万队列，消费延迟稳定在 12 毫秒，为百万级租户调用数十种大模型提供高隔离、低成本的流量保障。
阿里云 STAROps 助力企业打造业务数智运维体系
当业务系统与 AI 应用日益复杂，运维也需要从事后救火转向主动治理。阿里云全域智能运维平台 STAROps 以统一可观测数据为底座，通过 UModel 运维数字孪生还原应用、资源、告警和变更之间的关系，结合 AI 分析算子与运维智能体，形成“故障诊断—人工确认—自动处置—效果验证”的闭环。该体系已在塔斯汀、畅捷通、精臣等企业落地。其中，畅捷通的故障定位时间由平均 10 分钟以上缩短至 30 秒以内，验证了智能运维在复杂生产环境中的可用性与可控性。
阿里云 AgentLoop 支撑企业级智能体全生命周期观测与优化
智能体上线并非终点，如何发现质量退化、归因成本变化、拦截高风险变更，是规模化运营的新课题。阿里云智能体观测与优化平台 AgentLoop 以智能体运行轨迹为核心，连接观测审计、自动化评估与经验回灌，将推理、工具调用和数据访问从“黑盒”转化为全链路可追溯过程，并把真实业务轨迹沉淀为可回归的评估资产。在阿里云多智能体研发流水线中，团队在保持交付质量的前提下，将单次任务 Token 消耗降低23.4%，让智能体优化从人工试错走向数据驱动的持续进化。
从让智能体跑起来，到跑得稳、管得住、持续进化，四项获奖实践覆盖了 AI 应用从运行、调度、运维到优化的生产级全生命周期，它们既源于阿里云内部及客户真实场景，也已沉淀为可复制的产品能力。阿里云云原生应用平台将继续推进云原生基础设施面向 Agent 时代的创新升级，为企业智能化转型提供更可靠的技术底座。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
When artificial intelligence actually enters an enterprise ' s production environment, the measure of technology value is not just a model capacity, but also the ability to stabilize operations, fine movement, secure transport and continuous evolution. In recent days, the results of the Fifth Digital Transformation Call for the New Cup have been published, and the Alibaba Cloud cloud-based application platform team has been awarded four practical cases, covering AI agent operations, LLM traffic governance, digital communications peacekeeping and AI agent optimization throughout the life cycle, focusing on engineering capabilities that support the scaling down of enterprise-level AI applications.
The fifth Cup, hosted by the China Communications Standardization Association and officially launched by the China Institute of Information and Communication (CITI) in May 2026, lasted nearly five months and focused on innovations in the area of intellectual transformation of the country's economy, with the aim of providing the industry with replicable and replicable practical references.
Four awards for innovative practices covering the life cycle of AI agent
Qoder Cloud Ages large-scale AI agent hosting and operating practices based on RocketMQ LiteTopic
To “run up” AI agent on a large scale, first and foremost to resolve the contradiction between long-range missions, high-symmetry sessions and the cost of computing. Alibaba Cloud Qoder Cloud Argentinas build a cloud operating structure based on RocketMQ LiteTopic that is “brain separation, event-driven”: AI agent is responsible for decision-making, carrying out node tasks on demand and releasing resources upon completion; each session has an independent and orderly implementation channel, which can be followed by other nodes without loss. At present, the programme has supported the simultaneous distribution of all-class sessions, resulting in the “residentization of sessions, computation of mobility, continuity of status” and the extension of the message medium from a walk-through communication tool to the running base of AI agent.
RocketMQ LiteTopic supports the fine-tuned flow management of the LLM service platform
In the face of millions of class tenants calling LLM at the same time, the traditional rate limit can only decide whether to “leave or reject” but it is difficult to balance the smooth use of scarce calculus against the tenant experience. The Alibaba Cloud LLM service platform is based on RocketMQ LiteTopic, which constructs distributed leak drums with a "user+ model" dimension, and reprogramming rate limit into a fine rhythm schedule. Large-volume tenants share the pool, and the congestion of individual tenants does not affect other tenants. In production practice, single-message proxy nodes can carry 2 million queues, with consumption delays stabilizing at 12 milliseconds, providing high-segregation, low-cost flow security for millions of class tenants to call dozens of LLMes.
Alibaba Cloud STAROPs Helping Enterprises Build Business Counting Smart System
As operational systems and AI applications become more complex, the mobility also requires a shift from post-fire response to active governance. The Alibaba Cloud Global Smart Transport Platform STAROPs, based on the unified detectable data base, combines AI analysis algorithms with operational AI agent, through the relationship between UModel-dimensional digital-restorable applications, resources, alarms and changes, to form the "fail diagnosis-human confirmation-Auto-Disposal-Effect Verification" closed loop. The system has landed in enterprises such as Tasting, Quest and Quest. Among them, the fast-track failure-positioning time was reduced from an average of more than 10 minutes to less than 30 seconds, confirming the availability and control of intelligent transport in complex production environments.
Alibaba Cloud AgentLoop to support enterprise-level AI agent full-life observations and optimization
AI agent is not the end point, and how quality degradation is detected, attribution cost changes, and interception of high-risk changes are new topics for scale operations. The Alibaba Cloud AI agent Observatory and Optimization Platform AgentLoop, with the AI agent track at its core, connects observation audits, automated assessments and lessons learned, transforms reasoning, tool call and data access from a “black box” to a full chain traceability process, and settles real business tracks as reversible assessment assets. In the Alibaba Cloud excess AI agent development flow line, the team reduced single mission Token consumption by 23.4 per cent, while maintaining the quality of delivery, allowing AI agent to optimise from artificial error to continuous evolution driven by data.
From allowing AI agent to run smoothly, managed and continuously evolving, the four award-winning practices cover the full life cycle of AI applications from operation, movement, transport to optimized production levels, which are both internal to Alibaba Cloud and customer reality, and have sunk into replicable product capabilities. The Alibaba Cloud Cloud Native Applications Platform will continue to advance the upgrading of cloud native infrastructure towards the Agent era, providing a more reliable technological base for enterprise intellectual transformation.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.