随着 AI Agent 从简单的单轮问答走向复杂的编码与工具协作，任务执行周期被拉长至分钟甚至小时级。面对网络中断、资源争抢与状态丢失等生产级挑战，阿里云在 RocketMQ for AI 的持续演进中，于 9 月 22 日云栖大会 AI 实时数据智能论坛上正式发布 LiteTopic 的多项新特性，并展示了其在阿里云百炼与 Qoder Cloud Agents 中的落地实践。本文将深入解析如何利用 LiteTopic 构建高可靠的 Agent 异步通信架构。
趋势与挑战：当 Agent 遇上“长时任务”
当 AI Agent 开始承接编码、工具协作等复杂工作，一项任务可能经历多轮推理、外部执行与人工确认，持续数十分钟甚至更久。网络中断后能否接回已有输出，执行实例更换后能否延续进度，大量请求同时到来时如何分配算力，正在成为 Agent 进入生产环境必须解决的工程问题。
Agent 不再是简单的 Request-Response 模式，而是需要经历“提交-执行-等待-恢复-完成”的长生命周期。这种演进带来了三个核心挑战。首先是 “阻塞与等待”成为常态，复杂任务需要等待外部 Tool 回调、MCP 服务响应、Human-in-the-loop（人工确认），甚至排队等待 GPU 配额。其次是基础设施的不稳定性：在长达数小时的执行周期里，网络重连、Node 宕机、Pod 驱逐是不可避免的。第三是有限 GPU 资源的分配：需要在多用户之间公平调度有限算力，通过排队与错峰避免争抢和突发流量冲击。
针对长时任务场景，传统的内存态重试机制已难以胜任。为此，我们亟需一种全新的架构，通过将控制面与数据面解耦，来实现持久化执行（Durable Execution）。该架构能够将任务的状态与进度进行持久化存储，确保其可查询、可恢复，并支持实时通知。
业界共识：Agent Harness 的四个演进方向
为了实现上述目标，Agent 的执行控制面（Control Plane）正在发生深刻的演进。业界的核心共识已经凝聚在四个方向：状态外置、执行隔离、协议标准化以及协作可治理。
首先是架构解耦，走向 Brain-Hands-Session 的彻底分离。过去，组件往往绑定在同一个容器中，进程一旦崩溃，任务便全部丢失，且难以替换。如今，通过将控制面与数据面拆解，让大模型推理（Brain）、沙箱执行环境（Hands）和会话上下文（Session）作为独立组件协同工作。这使得沙箱实现了完全的无状态化，可以随时销毁与重建。
与此同时，原生支持持久化执行（Durable Execution）与异步人工协作（HITL），将任务状态进行 Checkpoint 持久化并转向事件驱动架构，以优雅应对网络故障、异步回调和人工审批等断点续传场景。
此外，通过采用 MCP 协议统一工具接入、利用 A2A 实现跨框架 Agent 互联，并引入 Agent Skills 按需动态加载上下文，有效缓解了 Context Window 的膨胀压力；最终在执行层从单体模式升级为分层协作架构——由主 Agent（Supervisor）专注意图拆解与结果验收，多个子 Agent（Worker）在独立上下文和工作空间中并行执行具体任务，彻底告别了低效的串行处理和逻辑迷失。
核心发布：RocketMQ LIteTopic 八大核心能力
为了解决上述问题，RocketMQ 推出了 LiteTopic 新特性。它将长时任务的“黑盒状态”转化为透明、可控的“消息事件流”。
RocketMQ 还通过标准协议与插件的方式，全面融入现有 Agent 生态架构，真正做到“少改业务代码，用标准协议与插件接入异步 Agent 通信”。
生产落地：通过 LiteTopic 实现会话状态的连续性
在生产环境中，Agent 与 LLM 的流式接口高度依赖 SSE/WebSocket 长连接，而网关滚动发布、Pod 重启、网络抖动等，都会导致长连接瞬间断开，已生成的 Token 无法送达、GPU 算力空转，用户被迫重发，成本与等待时间成倍增加。传统 Redis Stream 方案在此场景下同样捉襟见肘——TTL 从 Key 创建时计算而非消息写入时，不支持通配订阅导致父 Topic 无法一次性覆盖子会话，appendfsync 策略更在可靠性与性能之间难以兼顾。
针对这一痛点，RocketMQ LiteTopic 提供了“断连不丢、续连不重、体验无感”的会话连续性方案。其核心架构分为四层：多会话接入层接收来自 Web 端的并发连接，每个 Session 携带唯一 ID；无状态应用集群不依赖本地状态，任何节点均可接管任意 Session，支持横向扩展；LiteTopic 会话队列层为每个 Session 分配一条独立子队列（如 chat/Session1），归属统一的父 Topic CHAT_STREAM；大模型任务调度层按 SessionID 将流式 Token 写入对应 LiteTopic。当连接断开时，消费者只需携带上次消费位点（Offset）重新订阅，即可从断点处续推，无需重传已生成的 Token，真正实现了会话状态的连续性与成本的最优化。
生产落地：阿里云百炼通过 LiteTopic 实现细粒度隔离和限流
在共享 GPU 资源池中，平台需要同时考虑客户配额、基本服务机会与突发流量控制。若只使用全局限流，局部流量过高可能连带影响其他用户。
阿里云百炼的 LiteTopic 流程采用 userId 与 modelId 的组合进行路由，将不同用户对不同模型的请求写入相应队列。消费者通过父 Topic 的通配订阅，持续发现并处理动态产生的 LiteTopic。
限流组件按队列统计消费速率。当目标队列达到策略阈值时，通过 Suspend 暂停其消息投递，其他队列继续消费。条件恢复后，目标队列再按受控节奏处理积压请求。
这种方式将限流落实到用户与模型的具体组合，为配额保障和公平调度提供更细的执行手段。配合有界排队、等待时限和拒绝策略，平台可以缓冲短时超额请求，降低集中重试给下游带来的压力。实际可承接的流量仍取决于模型服务容量与调度策略。
生产落地：Qoder Cloud Agents 通过 LiteTopic 实现长程任务的状态托管与执行交接
Qoder Cloud Agents 的工程实践进一步展示了异步通信在复杂 Agent 运行时中的作用。其 Cloud Agent Runtime 面向 Serverless 云端 Agent 托管与执行，通过 LiteTopic 衔接任务步骤与执行结果。
针对长程任务，平台持久保存上下文与进度，为执行实例更换后的状态重建提供依据。针对大量并行会话，通过轻量业务通道组织相关消息，在满足通道内处理顺序要求的同时，让独立任务并行推进。
容量管理方面，模型与工具可以分池管理。对应资源暂时不足时，运行时局部暂停工作，并在条件具备、恢复事件到达后按需接手。等待期间能否释放具体资源，由运行时结合任务状态和执行环境决定。
通过这样的职责划分，平台承接长程任务的状态管理与执行交接，业务团队可以持续完善编码工作流，把开发重点放在任务完成质量与实际交付上。
生态集成与未来展望
在生态接入方面，RocketMQ 还通过标准协议与插件的方式，全面融入现有 Agent 生态架构，真正做到“少改业务代码，用标准协议与插件接入异步 Agent 通信”。
具体而言，RocketMQ 在生态适配层面实现了五大关键能力：
- MCP Agentic Messaging Primitives： RocketMQ 以 Tasks、Subscriptions/Listen 与 Progress Notifications 等原语，支撑长循环、流式结果输出与途中干预，未来线路图还将继续完善 Webhooks、Channels 等服务端事件能力，让大模型 Agent 的通信行为有据可依。
- A2A Transport 层面： RocketMQ 将传统的点对点 HTTP 协作升级为异步、可靠、可回溯的通信通道，天然支持跨集群协作、任务委派与结果回报，从根本上解决了 HTTP 同步调用在长时任务中的超时与不可靠问题。
- RocketMQ Flink Connector： 原生支持 LiteTopic 的消费。当 GPU 配额不足时，Flink worker 可以主动挂起消费，待资源恢复后从上次位点继续，避免消息积压持续冲击下游系统。
- OpenClaw RocketMQ Channel： 通过 LiteTopic 完成任务转发、结果汇聚与协同编排，开发者无需再自建消息通道，大幅降低了多 Agent 协同的工程成本。
- Spring Cloud Gateway RocketMQ Rate Limiter： 将传统的硬拒绝升级为 PASS、QUEUE、REJECT 三态缓冲：超限请求被异步卸载至 RocketMQ 队列，按受控速率逐步放行，仅在突发容量彻底耗尽时才返回 429，极大提升了系统的弹性与用户体验。
详细的代码见仓库：github.com/apache/rock…
随着 Agent 承担更长、更复杂的工作，任务能否在等待和故障之后继续推进，将直接影响用户体验与资源使用效率。RocketMQ for AI 围绕 LiteTopic 展开的实践，为会话续传、精细调度和云端 Agent 执行提供了具体的架构参考。欢迎加入 RocketMQ for AI 钉钉交流群，与产品和技术团队交流 LiteTopic 的能力演进、应用场景与落地实践。
钉钉搜索群号： 193120010825，立即加入 Kafka流算湖一体化交流群！
点击此处回看本场论坛直播。
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。
As AI Agent moves from simple single-cycle questions and answers to complex coding and tools, the task execution cycle is stretched to minutes and even hours. In the face of production-level challenges such as network disruptions, resource competition and loss of status, Alibaba Cloud, in the continuing evolution of RocketMQ for AI, presented several new features of the official publish LiteTopic at the AI Real-Time Data Smart Forum on September 22, and demonstrated its landing practices in Alibaba Cloud and Qoder Cloud Enterprises. This paper will provide an in-depth analysis of how LiteTopic can be used to construct a high-reliable Agent rectangular communication structure.
Trends and challenges: When Agent encounters a "long mission"
When AI Agent begins to take on complex tasks such as coding, tool collaboration, a mission may experience multiple rounds of reasoning, external execution and human confirmation, lasting several minutes or more. The possibility of retrieving existing outputs after network disruptions, the ability to sustain progress after replacement of implementation cases, and the distribution of computing power when a large number of requests arrive at the same time are becoming engineering issues that Agent has to resolve to enter the production environment.
Agent is no longer a simple Request-Response model, but needs to go through a long life cycle of “present-execut-wait-recover-complete”. This evolution presents three core challenges. The first is that “blocking and waiting” becomes the norm, with complex tasks waiting for external Tool, MCP service response, Human-in-the-load (human confirmation), and even waiting in line for GPU quotas. The second is the precariousness of infrastructure: in the several-hour implementation cycle, network reconnection, Node crash and Pod eviction are inevitable. The third is the allocation of limited GPU resources: there is a need for equitable movement of limited computing power among multiple users to avoid rivalry and sudden flow shocks by queuing and miscalculation.
The traditional internal retry mechanism has become incompetent for a long mission scenario. To that end, we urgently need a completely new structure to achieve sustainable implementation by decoupling control and data. The architecture allows the permanent storage of mission status and progress, ensures that it is searchable, recoverable and supports real-time notification.
Industry Consensus: Four Directions for the Evolution of Agent Harness
In order to achieve the above objectives, the executive control of Agent (Control Plane) is undergoing a profound evolution. The core consensus of industry has converged in four directions: external status, enforcement segregation, standardization of agreements, and collaboration to govern.
The first is to decorate the structure and move towards the complete separation of the Brain-Hands-session. In the past, components were often bound to the same container, and once the process collapsed, the task was completely lost and difficult to replace. Today, the LLM reasoning (Brain), Sandbox Implementation Environment (Hands) and Session context (Ssession) work as separate components by dismantling the control and data surfaces. This has rendered the sandbox completely non-stated and ready for destruction and reconstruction.
At the same time, the Durable Export works in collaboration with HITL to sustain the mission and shift to an event-driven structure to respond gracefully to disruptions such as network failure, walkbacks and manual clearances.
In addition, cross-frame Agent interconnection through the use of the MCP protocol integration tool, using A2A, and the introduction of Agent Skills to dynamically load context on demand, effectively eased the expansion pressure of Context Windows; eventually, the executive layer was upgraded from a monomer mode to a layered collaboration structure -- – Focused on dismantling and result acceptance check by Major Agent (Supervisor), and several sub-units of Agent (Worker) carry out specific tasks in parallel with independent context and workspaces, leaving out of inefficient serial and logic.
Core publish: RocketMQ LiteTopic
To solve these problems, RocketMQ introduced the LiteTopic new features. It transforms the “black box status” of the long-term mission into a transparent and manageable “message flow”.
RocketMQ is also fully integrated into the existing Agent ecological architecture by means of standard protocols and plugins, making it truly possible to "reduce business code changes and access different Agent communications with standard protocols and plugins".
Production landing: continuity of session status through LiteTopic
In the production environment, Agent 's current interface with LLM is highly dependent on the SSE/WebSocket long connection, while the gateway rolling publish, Pod restart, network shaking, etc., can lead to instant disconnections of long connections, generated Token is unable to reach, GPU algorithms are empty, users are forced to re-activate, and costs and waiting times multiply. The traditional Redis Stream scheme is equally stretched in this scenario - TTL calculates instead of writing messages from the time Key was created, does not support universal subscriptions that make it impossible for Father Topic to overwhelm sub-meetings, and appendfsync strategies are more difficult to balance between reliability and performance.
In response to this pain, RocketMQ LiteTopic offers a continuous programme of sessions “intermittent, continuous and unheard”. The core structure is divided into four tiers: multi-talking access layer receives a simultaneous connection from the Web end, each session carries a unique ID; a non-state application cluster does not depend on the local state, and any node can take over any session to support horizontal expansion; the LiteTopic session line assigns a separate queue (e.g. chat/session1) to each session, belonging to the unified parent Topic CHAT STREAM; and the LLM task schedule uses SesionID to write flow Token to the LieteTopic. When the connection is disconnected, the consumer can simply resubscribe the last consumption point (offfset) from the breakpoint without having to retransmit the generated Token, effectively optimizing the continuity and cost of the conversation.
Production landing: Alibaba Cloud calcium separators and rate limit by LiteTopic
In the shared GPU resource pool, the platform needs to consider both client quotas, access to essential services and sudden flow control. If only global rate limit is used, excessive local traffic may affect other users.
The Alibaba Cloud liteTopic process uses a combination of userId and modelId to route requests from different users for different models into the corresponding queue. Consumers continuously discover and process dynamic LiteTopic through parent Topic subscriptions.
rate limit components calculate consumption rates by queue. When the target queue reaches the strategic threshold, the delivery of messages is suspended through Suspend, and the others continue to consume. Once the conditions have been restored, the target queue will then process the backlog at the controlled rhythm.
In this way, rate limit is put into a specific mix of users and models, providing a more detailed means of implementation for quota guarantees and fair movement control. The platform, in conjunction with a queuing of boundaries, waiting time frames and a refusal strategy, can cushion short-term excess requests and reduce the pressure on downstream retry concentration. The actual available traffic continues to depend on model service capacity and movement strategy.
Production landing: Qoder Cloud Ages
The engineering practice of Qoder Cloud Agens further demonstrates the role of the different communication in complex Agent running times. Its Cloud Agent Runtme is directed towards Serverless cloud hosting and execution, using the LiteTopic bridging mission step and implementation result.
For long-range missions, the platform keeps context and progress on a permanent basis to provide a basis for the re-establishment of the status of the implementation case after replacement. In response to a large number of parallel sessions, information is organized through light business corridors, allowing for parallel independent missions while meeting the sequencing requirements within the corridors.
For capacity management, models and tools can be managed in pools. When counterpart resources are temporarily insufficient, operations are partially suspended and taken over as required when conditions are available and recovery events arrive. The ability to release specific resources during the waiting period is determined by the operational combination of mission status and implementation environment.
Through such a division of responsibilities, the Platform will be able to continuously refine the code workflow, focusing on the quality and actual delivery of tasks.
Ecological integration and future prospects
In terms of ecological access, RocketMQ is fully integrated into the existing Agent ecological architecture by means of standard protocols and plugins, making it truly possible to "reduce business code changes and communicate differently with a standard protocol and plugin."
Specifically, RocketMQ has achieved five key capabilities at the ecological adaptation level:
- MCP Agency Messaging Technologies: RocketMQ supports long cycle, flow-out and en route interventions in the original languages of Tasks, Subscriptions/Listen and Progress Studies, and will continue to refine the capabilities of service-end events such as WebHOKs, Channels and so on, to justify LLM Agent 's communications.
- At the A2A Transport level: RocketMQ upgrades the traditional point-to-point HTTP collaboration to a walk-in, reliable, traceable communication channel, natural support for cross-cluster collaboration, task assignment and return on results, fundamentally addressing the problem of HTTP synchronized calls for timeout and unreliability in long missions.
- RocketMQ Flink Contractor: Native support for LiteTopic consumption. When the GPU quota is insufficient, Flink worker can proactively hang up consumption, continue from the last position after resources are restored, and avoid a backlog of information continuing to shock downstream systems.
- OpenClaw RocketMQ Channel: Through LiteTopic, results are sent, results are assembled and co-organized, and developers are no longer required to build their own information channels, significantly reducing the cost of multiple Agent synergies.
- Spring Cloud Gateway RocketMQ Rate Limiteder: Upgrading traditional hard refusals to the PASS, QUEUE, REJECT Tri-state Buffer: the excess request was off-tracked to the RocketMQ queue, gradually released at a controlled rate, returning to 429 only when the surge capacity was completely depleted, greatly improving the system ' s elasticity and user experience.
For details, see repository: Github.com/apache/rock...
As Agent undertakes longer and more complex tasks, the ability of the mission to move forward after waiting and failure will have a direct impact on user experience and resource efficiency. The RocketMQ for AI practice around LiteTopic provides a specific framework reference for conversational continuity, fine scheduling and cloud implementation. Welcome to the RocketMQ for AI Nail Exchange, and share with product and technology teams the evolution of the LiteTopic capabilities, application scenes and landing practices.
Nailed search group: 193120010825, immediately join the Kafka flow count for lake integration!
Please click here to view the forum live.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.