RUNTIME v1.0ARGI Runtime now supports full ReAct loops, StateGraph orchestration, and real-time Studio trace debuggingExplore Architecture
Skip to main content
ARGI LogoARGIJAVA AGENT RUNTIME

ARGI

Agent Runtime and Graph IntelligencePronounced "AR-jee" · /ˈɑːr.dʒiː/

Production-grade Agent Runtime for Java builders. Build observable, recoverable, and collaborative multi-agent systems with ReAct loops, StateGraph orchestration, durable checkpoints, and context engineering.

ReAct Reasoning LoopStateGraph OrchestrationDurable Checkpoints & HITLMCP & Multi-Agent A2A
JAVA //ReactAgent.builder().model(chatModel).tools(mcpTools).saver(checkpointSaver).build()
AGENTIC_RUNTIME_TELEMETRY
ckpt_react_904a
STARTPLANReAct LLMROUTEConditionACTMCP ToolVERIFYGuardrailSTATECheckpointOUTPUT
MODE // REACT_LOOP42ms TTFT
ReAct Autonomous Reasoning LoopLIVE EXECUTION TRACE
THINK
ReactAgent.plannerDecompose query into multi-step retrieval + verification plan
38ms
ACT
McpToolNode.invokeExecute sandboxed MCP tool `queryClusterMetrics(region="cn-hangzhou")`
64ms
OBSERVE
ContextEngine.mergeAppend structured tool observation & compress sliding context window
12ms
CHECKPOINT
StateSaver.commitPersist state snapshot #ckpt_react_904a for zero-loss recovery
4ms
CODE-TO-RUNTIME BLUEPRINTS // 声明式智能体工程

From Declarative Java to Autonomous Graph Execution

Move beyond fragile prompt chains and opaque loops. Gain durable state checkpoints, parallel fan-out/fan-in branches, and human-in-the-loop governance on top of standard Java & Spring.

ProductionReactAgent.java
ReactAgent opsAgent = ReactAgent.builder()
    .name("cloud-sre-agent")
    .model(chatModel)
    .systemPrompt("You are an autonomous production SRE agent.")
    .tools(
        mcpClient.resolve("prometheus-query"),
        mcpClient.resolve("k8s-rollout-status")
    )
    .hooks(
        ModelCallLimitHook.of(12),
        ContextCompressionHook.slidingWindow(16_384)
    )
    .saver(PostgresCheckpointSaver.create(dataSource))
    .build();

AssistantMessage reply = opsAgent.call(
    "Diagnose latency spike on payment-service and propose safe rollback",
    RunnableConfig.builder().threadId("incident-2026-09").build()
);
DECLARATIVE AGENT BUILDER

几行 Java 代码构建生产级 ReAct 智能体

通过流式 Builder API 将大模型推理、MCP 工具集、上下文压缩钩子与持久化检查点组合为可观测的自治智能体循环。

RUNTIME STATE SNAPSHOTPERSISTED (PostgreSQL)
thread_id: incident-2026-09node: ReactAgent.toolExecution
agent.name"cloud-sre-agent"
loop.iteration3 / 12 (CONVERGED)
mcp.tools.invoked["prometheus-query", "k8s-rollout-status"]
context.tokens4,820 / 16,384 (COMPRESSED)
  • 原生支持 MCP 工具协议与 Spring AI ToolCallback 自动装配
  • 内置 Hook 生命周期拦截(模型调用限流、PII 脱敏、上下文压缩)
  • 每次推理与工具调用自动生成可回放的状态快照
RUNTIME ARCHITECTURE // 核心技术底座

Upper-layer agent design, runtime-grade execution

Graph 和 ReAct Agent 承担编排、状态、恢复和协作语义,让复杂 Agent 应用具备可观察、可恢复和可扩展的运行时边界。

CORE // 01AUTONOMOUS LOOP

ReAct Agent 自治推理引擎

将思考(Reason)、工具行动(Act)与环境反馈(Observe)组织成受控的闭环执行流,内置模型限流、上下文压缩与结构化输出校验。

探索 ReAct Agent →
GRAPH // 02DAG + CYCLIC

Graph Core 状态图编排

用强类型 OverAllState、条件边(Conditional Edges)、并行分支与嵌套子图描述长周期复杂工作流,兼顾确定性控制与动态智能路由。

探索 Graph Core →
STATE // 03TIME-TRAVEL

持久化检查点与故障自愈

每个节点边界自动持久化状态快照(Memory / Redis / PostgreSQL),支持跨实例故障恢复、状态分叉(Fork)与历史回放。

了解持久化执行 →
GOVERNANCE // 04HUMAN-IN-LOOP

人在回路(HITL)安全管控

支持在敏感工具或高危节点执行前声明式中断(interruptBefore),人工审批或热修改状态参数后无缝恢复原执行上下文。

查看 HITL 模式 →
STUDIO // 05VISUAL TRACE

Studio 嵌入式可视化调试

内嵌交互式调试工作台,实时渲染智能体对话流、StateGraph 拓扑高亮、工具入参/出参及节点级耗时遥测。

启动 Studio 调试 →
MESH // 06MCP + A2A

MCP 协议与多智能体协同网络

原生集成 Model Context Protocol(MCP)工具生态,支持 A2A 远程 Agent 封装,以及 Sequential、Parallel、LlmRouting、Loop 与 Agent-as-tool 编排。

探索多智能体编排 →
MODULAR STACK // 全栈智能体模块矩阵

ARGI Ecosystem

A comprehensive ecosystem for building intelligent applications

LAYER 01 // AGENT RUNTIME

ARGI ReAct Agent

Upper-layer framework for ReAct Agent, stateful agent loops, and multi-agent coordination.

ReAct LoopHooks & GuardrailsContext CompressionMulti-Agent
LAYER 02 // STATEFUL GRAPH

ARGI Graph Core

Graph runtime for workflow orchestration, state checkpoints, recovery, and human-in-the-loop execution.

StateGraphConditional EdgesDurable CheckpointsHITL
LAYER 03 // VISUAL DEBUGGER

ARGI Studio

Embedded visual debugging studio for agent conversations and graph workflows.

Visual DAGTrace ReplayState MutationToken Telemetry
LAYER 04 // PROTOCOL & MESH

ARGI Extensions

Pluggable ecosystem for model adapters, distributed persistence backends, A2A discovery, MCP registry, and sandboxed tool execution.

MCP RegistryA2A ProtocolRedis / Postgres SaverSandbox

ARGI(Agent Runtime and Graph Intelligence)是面向 Java 开发者的智能体运行时与工作流框架。