Spring AI Integrations

langgraph4j-springai-agentexecutor provides ReAct-style agents built on LangGraph4j and Spring AI ChatModel. Use it when you want a ready-to-run agent loop with tool callbacks, streaming output, approvals, or sub-agents instead of wiring every node manually.

Features

AgentExecutor for a compact ReAct graph with agent -> action -> agent execution.
AgentExecutorEx for explicit tool-dispatch nodes, approval gates, and richer orchestration.
Streaming support backed by Spring AI ChatModel responses.
Tool registration from ToolCallback, ToolCallbackProvider, or annotated tool objects.
Optional LangGraph Studio integration for interactive graph execution.

Agent Executor Diagram

flowchart TD
    __START__((start))
    __END__((stop))
    agent("agent")
    action("actions")
    %%  condition1{"check state"}
    __START__:::__START__ --> agent:::agent
    %%  agent:::agent -.-> condition1:::condition1
    %%  condition1:::condition1 -.->|continue| action:::action
    agent:::agent -.->|continue| action:::action
    %%  condition1:::condition1 -.->|end| __END__:::__END__
    agent:::agent -.->|end| __END__:::__END__
    action:::action --> agent:::agent

    classDef ___START__ fill:black,stroke-width:1px,font-size:xx-small;
    classDef ___END__ fill:black,stroke-width:1px,font-size:xx-small;

AgentExecutorEx Diagram

flowchart TD
    __START__((start)):::__START__
    __END__((stop)):::__END__
    model("model")
    action_dispatcher("action_dispatcher")
    action1("action 1")
    action2("action 2")
    approval_action3("approval action 3 <br>(interruption)")
    action3("action 3")

    %%  condition1{"check state"}
    %%  condition2{"check state"}
    __START__:::__START__ --> model:::model
    %%  model:::model -.-> condition1:::condition1
    %%  condition1:::condition1 -.->|continue| action_dispatcher:::action_dispatcher
    model:::model -.->|continue| action_dispatcher:::action_dispatcher
    %%  condition1:::condition1 -.->|end| __END_:::__END_
    model:::model -.->|end| __END__:::__END__
    action1:::action1 --> action_dispatcher:::action_dispatcher
    action2:::action2 --> action_dispatcher:::action_dispatcher
    %%  action_dispatcher:::action_dispatcher -.-> condition2:::condition2
    %%  condition2:::condition2 -.-> model:::model
    action_dispatcher:::action_dispatcher -.-> model:::model
    %%  condition2:::condition2 -.-> __END_:::__END_
    action_dispatcher:::action_dispatcher -.-> __END__:::__END__
    %%  condition2:::condition2 -.-> action1:::action1
    action_dispatcher:::action_dispatcher -.-> action1:::action1
    %%  condition2:::condition2 -.-> action2:::action2
    action_dispatcher:::action_dispatcher -.-> action2:::action2
    %%  condition1{"check state"}
    %%  condition2{"check state"}

        action3:::action3 --> action_dispatcher:::action_dispatcher
    approval_action3:::approval_action3 -.-> model:::model
    approval_action3:::approval_action3 -.-> action_dispatcher:::action_dispatcher
    approval_action3:::approval_action3 -.->|APPROVED| action3:::action3
    action_dispatcher:::action_dispatcher -.-> approval_action3:::approval_action3


    classDef __START__ fill:black,stroke-width:1px,font-size:xx-small;
    classDef __END__ fill:black,stroke-width:1px,font-size:xx-small;

Installation

<dependency>
    <groupId>org.bsc.langgraph4j</groupId>
    <artifactId>langgraph4j-springai-agentexecutor</artifactId>
    <version>1.9.0</version>
</dependency>

This module depends on langgraph4j-spring-ai and targets Java 17.

Usage

Configure a `ChatModel`

@Configuration
public class ChatModelConfiguration {

    @Bean
    @Profile("ollama")
    ChatModel ollamaModel() {
        return OllamaChatModel.builder()
                .ollamaApi(new OllamaApi("http://localhost:11434"))
                .defaultOptions(OllamaOptions.builder()
                        .model("qwen2.5:7b")
                        .temperature(0.1)
                        .build())
                .build();
    }
}

Build and run an agent

var agent = AgentExecutor.builder()
        .chatModel(chatModel)
        .tools(tools)
        .build()
        .compile();

var result = agent.stream(
        GraphInput.args(Map.of("messages", new UserMessage("what is 234 + 45?"))),
        RunnableConfig.empty());

var finalState = result.stream()
        .reduce((a, b) -> b)
        .orElseThrow()
        .state();

The default state type is AgentExecutor.State, which extends MessagesState<Message>.

Enable streaming and tool extraction from an object

var agent = AgentExecutor.builder()
        .chatModel(chatModel)
        .streaming(true)
        .emitStreamingEnd(true)
        .toolsFromObject(new TestTools())
        .build()
        .compile(compileConfig);

Use `AgentExecutorEx` when you need approvals or sub-agents

var agent = AgentExecutorEx.builder()
        .chatModel(chatModel)
        .streaming(true)
        .approvalOn("threadCount", (nodeId, state) ->
                InterruptionMetadata.builder(nodeId, state)
                        .addMetadata("label", "confirm thread count execution?")
                        .build())
        .toolsFromObject(new TestTools())
        .build(compileConfig);

AgentExecutorEx.State adds channels for pending tool execution requests, tool responses, and next-action dispatching. This is the variant used by the test applications when approval flows or sub-agents are involved.

LangGraph Studio

The test configuration in src/test/java/.../LangGraphStudioConfiguration.java shows how to expose an AgentExecutorEx graph through LangGraphStudioConfig, backed by a MemorySaver checkpoint store.