How Apache RocketMQ Powers Asynchronous Multi‑Agent AI Architectures

In a keynote at the 2025 Global Machine Learning Conference, Alibaba Cloud senior technical expert Zhou Li presented "Apache RocketMQ x AI: Event‑Driven Architecture for Asynchronous Agents," describing how RocketMQ’s latest capabilities can build asynchronous multi‑agent systems.

Agentic AI and Multi‑Agent Architecture

With larger models and lower inference costs, AI has entered the Agentic AI era, shifting from passive response to proactive decision‑making. Multi‑Agent architectures distribute tasks among specialized agents, balancing model autonomy with business controllability. Two key challenges are agent ability discovery and task‑loop closure.

1. Agent Ability Discovery

Agents must dynamically register capabilities (e.g., "I can perform data analysis") and allow a Supervisor agent to query and select suitable Sub‑Agents at runtime, similar to service discovery in micro‑services but driven by semantic intent processed by LLMs.

2. Task Coordination

Effective coordination requires a communication mechanism. Three patterns are discussed:

Polling communication : simple but high latency and resource‑intensive.

Point‑to‑point invocation : explicit calls (REST, gRPC) with strong consistency but tight coupling.

Publish‑subscribe (Pub/Sub) : decoupled, scalable, and suitable for asynchronous workflows.

For asynchronous Pub/Sub, feedback from Sub‑Agents to the Supervisor can be handled via exclusive queues, broadcast filtering, or shared storage, each with trade‑offs.

RocketMQ Innovations for Agentic AI

RocketMQ introduces two features to address these challenges:

Semantic Topic : Topics become carriers of business intent and capability semantics, enriched with natural‑language descriptions and metadata, enabling agents to discover and bind to appropriate topics.

Lite‑Topic : A lightweight, dynamically created topic that does not require pre‑registration, supporting short‑lived, high‑frequency messages and automatic lifecycle management, ideal for result callbacks.

Semantic Topic as Intent Carrier

Each topic now includes a natural‑language description and structured schema, allowing a Supervisor to query the NameServer for topics matching required capabilities (e.g., "data analysis"), and dynamically route tasks to suitable Sub‑Agents.

Lite‑Topic and Subscription Management

Subscriptions are managed in a decentralized, eventually consistent manner using an InterestSet per client. Brokers store these interest sets and, upon new messages, populate a ReadySet that lists only active topics ready for consumption, turning traditional pull into a push‑enabled pull model.

End‑to‑End Asynchronous Multi‑Agent Workflow

Ability registration : Sub‑Agents create semantic topics and register schemas with the NameServer.

Semantic‑driven orchestration : The Supervisor builds prompts, queries available topics, and treats them as callable functions for the LLM.

Lightweight async dispatch : Tasks are sent to target topics; a temporary Lite‑Topic is created for the result callback.

Closed‑loop decision making : Supervisors subscribe to the Lite‑Topic, aggregate results, inject feedback, and trigger the next reasoning round.

This architecture leverages RocketMQ’s extended Pub/Sub model to provide loosely coupled, highly scalable, and low‑latency communication for multi‑agent AI systems, supporting dynamic task orchestration, result feedback, and iterative decision loops.