Spring AI Alibaba is a Java framework for building AI agents. Chapter 3 of the AI 原生应用架构白皮书 (AI-native application architecture whitepaper) hosted on jimmysong.io uses it as its worked example of agent development, observing that the agent development ecosystem began mainly in Python and later spread to languages such as Java, and presenting Java frameworks such as Spring AI Alibaba as giving developers extensive agent-building capabilities.
That chapter uses the framework to illustrate three development paradigms: a single agent, which adds retrieval, tools and memory to an LLM application (the chapter calls this an Augmented LLM Application — see Augmented LLM); a workflow, in which sub-agents are orchestrated along a predefined flow; and a multi-agent system, in which the flow between agents is decided by the model and the sub-agents collaborate with more autonomy (see LLM-Based Multi-Agent System).
Capabilities
The framework’s agent is defined as a ReactAgent, a typical implementation of the ReAct pattern
(see ReAct): a loop of thought, in which the agent analyses the goal, what it
knows and which tools are available; action, in which it calls an external tool with a structured
instruction; and observation, in which the tool’s result is fed back — repeated until the task is
complete. Defining an agent, as the chapter describes it, takes a name used as a unique identifier, a
description of its capabilities so that agents can be told apart in multi-agent settings, and a model,
with “qwen-max” given as an example. An instruction parameter guides its behaviour — the core task or
goal, a persona, behavioural constraints, instructions for using tools and the required output format —
and tools can be supplied as native functions, APIs or other agent instances.
The chapter’s comparison of the framework’s built-in agent types is:
| Type | Core behaviour | Flow orchestration | Determinism |
|---|---|---|---|
| ReactAgent | Reasoning, generation, tool use | None | Relatively high uncertainty |
| FlowAgent | Controls the flow of several agents | Predefined flow (sequential, parallel, etc.) | Relatively high determinism |
| MultiAgent | Controls the flow of several agents | Flow control that supports model decisions | Between ReactAgent and FlowAgent |
For workflows it describes three agent types. SequentialAgent runs sub-agents one after another, each one’s output becoming the next one’s input; the example is a writing assistant made of a writer agent and a reviewer agent. ParallelAgent runs sub-agents in parallel, suited to tasks spanning several domains; the example is a search assistant whose domain-specific sub-agents run in parallel, with a merger agent combining their output and the whole chained together with a SequentialAgent. LoopAgent runs sub-agents repeatedly for tasks that need several rounds of iteration. For model-driven multi-agent systems, LlmRoutingAgent lets the model decide which sub-agent the flow goes to next, for dynamic flow control.
Adoption & Ecosystem
The same chapter goes on to cover moving agents from a single process to distributed deployment through the Agent2Agent Protocol (A2A), with Nacos acting as a registry for agents, and message-driven communication between agents built on RocketMQ. It presents these as the next stage after the in-process patterns above rather than as features of the framework itself.