Ragent: Ant Group’s Ray‑Based Distributed Agent Framework

Ant Group presents Ragent, a new distributed agent framework built on Ray, aiming to support large‑model AI workloads at massive scale.

Background : Ray, originally developed by OpenAI for large‑model training, has been adopted by Ant Group since 2017, contributing over 26% of Ray’s core code and operating more than 1.5 million CPU cores.

Motivation : Traditional service‑oriented applications struggle with the rapid prototyping, heterogeneous resource requirements, and complex deployment pipelines of AI agents; a unified, scalable solution is needed.

Design & Implementation : Ragent sits atop Ray and Kubernetes, providing a multi‑layer architecture: business layer (Agent Apps), Agent Crafting Platform, algorithm libraries (e.g., LangChain, MetaGPT), and the Ragent SDK that handles distributed execution, memory, tooling, and failover. It leverages Ray primitives (Task, Actor, Object) and Ray Data for batch processing.

The framework defines core agent modules—Profile, Memory, Planning, and Action—and supports tool registration via annotations, enabling agents to invoke external services such as LlamaIndex, Lucene, or custom DB queries.

Ragent also supports multi‑agent workflows, exemplified by a MetaGPT‑style software development pipeline where agents act as product manager, architect, coder, and tester, coordinated through an environment component that tracks tasks and message queues.

Future work includes developing an Agent Mesh protocol for standardized communication across different agent frameworks and extending support for diverse hardware accelerators.