Spring AI Alibaba

The project has three layers. At the bottom, Spring AI supplies model, tool, MCP, message and vector-store abstractions. Above it, Spring AI Alibaba Graph is a state-graph runtime with nodes, conditional edges, subgraphs and parallel branches, plus persistence and streaming. The Agent Framework sits on top: a ReactAgent runs a model-and-tool loop, and flow agents compose several agents.

Multi-agent setups come in two styles. SequentialAgent, ParallelAgent, RoutingAgent / LlmRoutingAgent and LoopAgent run sub-agents in a fixed or model-chosen order, passing data through shared state: each agent writes to an outputKey, and later instructions reference it with placeholders such as {article}. The docs also describe agents used as tools by a supervisor, and handoffs where the active agent changes. For full control, developers can build the flow directly with the Graph API.

State is an OverAllState map persisted by checkpoint savers, from in-memory to Redis, MongoDB or relational databases. Checkpoints make human-in-the-loop pauses and long-running tasks resumable. Conversation memory follows Spring AI's chat memory repositories, and the A2A starter integrates with Nacos so agents in different services can find and call each other.

Human review is a hook. A HumanInTheLoopHook is configured with approvalOn(toolName, ToolConfig) for each tool that needs oversight and added to the agent's hooks. When the model proposes one of those calls, the hook raises an interrupt, the graph state is saved by the checkpoint saver (for example MemorySaver), and execution stops. A person then approves the call as is, edits its arguments before it runs, or rejects it with an explanation that is added to the conversation. When several calls pause together, each gets its own decision. The Graph Core docs also include human-in-the-loop, cancellation and time-travel examples at the graph level.

<dependency>
  <groupId>com.alibaba.cloud.ai</groupId>
  <artifactId>spring-ai-alibaba-agent-framework</artifactId>
  <version>1.1.2.0</version>
</dependency>

import com.alibaba.cloud.ai.graph.OverAllState;
import com.alibaba.cloud.ai.graph.agent.ReactAgent;
import com.alibaba.cloud.ai.graph.agent.flow.agent.SequentialAgent;
import java.util.List;
import java.util.Optional;

// chatModel: any Spring AI ChatModel bean, e.g. from the DashScope or OpenAI starter
ReactAgent writer = ReactAgent.builder().name("writer").model(chatModel)
    .instruction("Write a 100-word product note about: {input}").outputKey("draft").build();
ReactAgent editor = ReactAgent.builder().name("editor").model(chatModel)
    .instruction("Tighten this note and return only the text: {draft}").outputKey("final").build();

SequentialAgent pipeline = SequentialAgent.builder()
    .name("note_pipeline").subAgents(List.of(writer, editor)).build();

Optional<OverAllState> result = pipeline.invoke("a reusable water bottle");
result.flatMap(s -> s.value("final")).ifPresent(System.out::println);

• Requires JDK 17+ (and Maven 3.8+ per the quick start).
• A model starter is separate: the docs use spring-ai-alibaba-starter-dashscope with AI_DASHSCOPE_API_KEY; other providers need the matching Spring AI starter.
• The quick start pins spring-ai-alibaba-agent-framework 1.1.2.0, while GitHub's latest stable release is v1.1.2.2 (2026-03-10).
• v2.0.0-M1.1 (2026-06-25) is a pre-release milestone that moves to Spring AI 2.0.0-M1 and Spring Boot 4.0.0.
• Human-in-the-loop needs a checkpoint saver configured on the agent; MemorySaver does not survive restarts.

Yes, through Spring AI. The docs show a ReactAgent using remote MCP tools via spring-ai-starter-mcp-client over streamable HTTP or SSE, and the Admin platform includes MCP management.

Yes. It can expose a ReactAgent as an A2A server and call remote agents with A2aRemoteAgent, with Nacos as an optional registry for agent cards.

Yes. It is Apache-2.0 licensed; model usage is billed by whichever provider you configure, such as DashScope or OpenAI.

JDK 17 or newer. The stable 1.1.x line builds on Spring AI 1.1 and Spring Boot 3.5; the 2.0 milestone moves to Spring AI 2.0.0-M1 and Spring Boot 4.

The repository was pushed on 2026-09-16. The overview says ReactAgent will get fixes and security patches while the team focuses on Spring AI integration and multi-agent work, pointing to AgentScope for a more advanced ReactAgent.