一句话：提出会话中心的可重放事件流模型与 Lite-Topic，解决大规模多智能体协作的突发流量、隔离及故障恢复难题。
论文研究阿里云生产级多智能体系统（MAS）的部署瓶颈，核心挑战在运维而非模型：突发流量导致进程内排队、内存压力与 GC 抖动，海量并发下会话隔离昂贵，中间上下文未持久化使故障难恢复。作者提出 RocketMQ-A2A，将每次协作会话建模为持久、可重放的事件流，并用消息中间件解耦任务调度与结果投递，为数千万并发会话提供独立回复流、重放恢复和可审计 A2A 调用链；同时开源 Lite-Topic，以 TTL 管理低开销短生命周期回复通道。评测显示，持久 MQ 积压吸收突发压力，在 25 倍负载下降低堆增长与 Full GC；Lite-Topic 支持数千万级会话隔离；重放使中断任务在下游故障后恢复完成。系统已部署于阿里云百炼。
We investigated deployment bottlenecks of production multi-agent systems (MAS) at Alibaba Cloud. We found that the dominant challenges are operational rather than model-centric: bursty traffic causes in-process queuing, memory pressure, and garbage-collection (GC) jitter; session isolation becomes expensive under massive concurrency; and failures are hard to recover from when intermediate context is not persistently recorded and replayable. We propose RocketMQ-A2A, a session-centric interaction model for production MAS. RocketMQ-A2A models each collaboration session as a persistent, replayable event stream, and decouples task scheduling from result delivery through message middleware. This design targets up to tens of millions of concurrent sessions, each with an independent reply stream, and provides replay-based recovery and auditable traces for asynchronous Agent-to-Agent (A2A) call chains without requiring application teams to implement custom isolation, ordering, and replay mechanisms. We implemented and open-sourced RocketMQ-A2A and Lite-Topic in Apache RocketMQ. Lite-Topic provides low-overhead, short-lived session reply channels with TTL-based lifecycle management. Our evaluation shows that durable MQ backlogs absorb burst pressure and reduce downstream heap growth and Full GC under 25× load, Lite-Topic supports session isolation at the tens-of-millions scale, and replay enables interrupted tasks to resume and complete after downstream failures. Our solution has been deployed in production on Alibaba Cloud Model Studio (Bailian), an enterprise-grade platform for building and operating LLM applications end-to-end.
@inproceedings{ZhouZJZZFL26,
title = {RocketMQ-A2A: Reliable Session-Level Replayable Event Streams for Large-Scale Multi-Agent Collaboration},
author = {Li Zhou and Shuo Zhang and Juntao Ji and Shijie Zhang and Ke Zhao and Yubao Fu and Qingshan Lin},
booktitle = {FSE},
year = {2026},
doi = {10.1145/3803437.3805231}
}