免费领取大会全套演讲PPT    

点击领取

我要参会

Wenting Yang

Product Expert, Alibaba Cloud

Wenting Yang is a Product Expert for Alibaba Cloud’s Intelligent Messaging Service RocketMQ, where she is responsible for product planning and implementation of RocketMQ for AI Agent scenarios. She has led the transformation of RocketMQ from a traditional messaging middleware into an AI-native asynchronous communication engine, and designed and drove the large-scale adoption of the LiteTopic lightweight messaging model across large-model platforms such as Alibaba Cloud Model Studio (Bailian). Her work addresses key communication challenges in Agent systems, including long-session persistence, multi-tenant isolation and rate limiting, and asynchronous Multi-Agent collaboration. She closely follows the evolution of the MCP protocol and emerging industry architectures such as Anthropic Managed Agents, with a focus on deeply integrating messaging capabilities with Agent communication standards. She is also an active contributor to the Apache RocketMQ open-source community.

Topic

#Rebuilding the Foundation for Agent Asynchronous Communication: Apache RocketMQ in Practice Through the Evolution of MCP and the Managed Agents Architecture

The MCP 2026 roadmap identifies Transport Scalability and Agent Communication as top priorities, while Anthropic’s Managed Agents architecture introduces a new paradigm built around Brain–Hands decoupling, persistent session logs, and event-driven communication. Together, these industry signals point to a clear conclusion: the Agent ecosystem is creating demand for an entirely new layer of asynchronous communication infrastructure. Starting from the tension between stateful sessions and horizontal scalability in the MCP protocol, as well as the append-only event log and failure recovery patterns of Managed Agents, this talk explores how RocketMQ addresses these emerging requirements through LiteTopic, which supports tens of millions of lightweight Topics, and Consumer Group Label, which enables fine-grained isolation and routing. Drawing on production validation from Alibaba Cloud Model Studio, the session demonstrates how Apache RocketMQ can evolve from a passive application messaging pipeline into an Agent Communication Operating System. Outline 1. Understanding Industry Trends: What Do the MCP 2026 Roadmap and Anthropic’s Managed Agents Architecture Reveal? The core tension in MCP: Streamable HTTP enables MCP Servers to become remote and distributed, but stateful Sessions inherently conflict with load balancing and horizontal scaling. The design philosophy of Managed Agents: Brain–Hands decoupling, stateless containers, and Sessions maintained as append-only persistent logs. A shared direction: Agents need a layer of stateful asynchronous communication infrastructure that is reliable, recoverable, and scalable. 2. Why the Agent Ecosystem Requires a New Definition of Message Middleware Lifecycle gaps in the MCP Tasks primitive—including retry semantics and result expiration policies—are fundamentally problems of reliable message delivery. The emitEvent/getEvents pattern in Managed Agents is essentially a model for producing and consuming ordered message streams. Traditional Topic models face challenges in both cost and granularity when supporting massive numbers of lightweight channels under a “one channel per Session” model. 3. Rethinking RocketMQ: How LiteTopic Addresses Agent Communication Requirements The design decision: supporting tens of millions of Topics enables a “one channel per Session” architecture, externalizing stateful MCP Sessions into persistent message streams. Mapping to the Managed Agents architecture: LiteTopic serves as a persistent Session Log, while message offsets correspond to the positional slices used by getEvents. Brain failure recovery: after an Agent restarts, it can resume consumption from the last message offset, naturally implementing wake(sessionId) semantics. 4. Production Scenario I: Persistent Recovery for Long-Lived MCP Sessions The challenge on Alibaba Cloud Model Studio: WebSocket or SSE disconnections can result in lost Session context. The solution: each MCP Session corresponds to a LiteTopic, with persistent messages enabling distributed failure recovery. The result: MCP Servers can achieve truly stateless horizontal scaling, making Session migration transparent to users. 5. Production Scenario II: Fine-Grained Isolation and Rate Limiting for Agent Requests The “Many Brains, Many Hands” model of Managed Agents requires flexible message routing. The solution: Consumer Group Label enables message isolation and canary releases for different Agents under the same Topic. Mapping to future MCP directions: an event-driven implementation path for features such as Triggers and Event-driven Updates. 6. Production Scenario III: GPU Resource Utilization and Asynchronous Orchestration The batchReceive consumption model aligns with the elastic architecture of Managed Agents, where containers are launched on demand. Asynchronous messaging replaces synchronous call chains, helping eliminate cascading timeouts across Agents. 7. Product Insights and Outlook The evolution of message middleware from an “application-to-application pipeline” to an “Agent Communication Operating System.” Opportunities for standardized integration of messaging capabilities into the future evolution of the MCP protocol. A new product thinking framework for infrastructure software product managers building for the Agent ecosystem. Audience Takeaways Frontier Perspective Gain firsthand insight into the MCP 2026 roadmap and Anthropic’s Managed Agents architecture, understand the evolution of Agent communication infrastructure, and develop an analytical framework for identifying product opportunities through protocol and standards trends. Product Thinking Upgrade Understand why the Agent ecosystem is redefining foundational software products. Message middleware is no longer an optional performance optimization component—it is becoming an essential communication layer for reliable Agent operation. Learn how to derive product requirements from emerging architectural and industry trends. Production-Proven Practices Explore three solutions validated at scale on Alibaba Cloud Model Studio: MCP Session persistence, multi-Agent isolation and routing, and elastic GPU scheduling. These approaches can be directly referenced when designing communication architectures for AI platforms. Future Trend Assessment Understand the relationship between emerging MCP capabilities—including Tasks, Triggers, and Event-driven Updates—and messaging system capabilities, helping you anticipate and position for the standardization of Agent communication infrastructure.

© boolan.com 博览 版权所有

沪ICP备15014563号-6

沪公网安备31011502003949号