中文主页

An agent-oriented programming framework for building LLM applications in Java.

AgentScope provides a comprehensive toolkit for creating intelligent agents with tool calling, memory management, multi-agent collaboration, and more.

Easy for beginners, powerful for experts.

• Transparent to Developers: Transparency is our FIRST principle. Prompt engineering, API invocation, agent building, workflow orchestration - all visible and controllable. No deep encapsulation or implicit magic.
• Realtime Steering: Native support for realtime interruption and customized handling.
• More Agentic: Support agentic tools management, agentic long-term memory control and agentic RAG, etc.
• Model Agnostic: Programming once, run with all models (DashScope, OpenAI, Anthropic, and more).
• LEGO-style Agent Building: All components are modular and independent.
• Multi-Agent Oriented: Designed for multi-agent, explicit message passing and workflow orchestration with Pipeline support.
• Reactive Architecture: Built on Project Reactor for efficient non-blocking async operations.
• Multimodal Support: Native support for vision, audio, and video content processing.
• Highly Customizable: Tools, prompt, agent, workflow, hooks, and visualization - customization is encouraged everywhere.

AgentScope Java requires JDK 17 or higher.

<dependency>
    <groupId>io.agentscope</groupId>
    <artifactId>agentscope</artifactId>
    <version>1.0.0</version>
</dependency>

Start with a basic ReActAgent that replies to user queries!

public static void main(String[] args) {
	ReActAgent agent = ReActAgent.builder()
			.name("Assistant")
			.sysPrompt("You are a helpful AI assistant.")
			.model(DashScopeChatModel.builder()
					.apiKey(System.getenv("DASHSCOPE_API_KEY"))
					.modelName("qwen-max")
					.build())
			.build();

	Msg userMessage = Msg.builder()
			.textContent("Hello, please introduce yourself.")
			.build();

	Msg response = agent.call(userMessage).block();
	System.out.println("Agent Response: " + response.getTextContent());

1. Define Tool

   Define a tool class with methods annotated with @Tool. Here's an example SimpleTools class with a time tool:

   public class SimpleTools {
       @Tool(name = "get_time", description = "Get current time string of a time zone")
       public String getTime(
               @ToolParam(name = "zone", description = "Time zone, e.g., Beijing") String zone) {
           LocalDateTime now = LocalDateTime.now();
           return now.format(DateTimeFormatter.ofPattern("yyyy-MM-dd HH:mm:ss"));
       }
   }

2. Register Tool to ReActAgent

   Register the tool class through Toolkit using the registerTool method:

   public static void main(String[] args) {
       Model model = DashScopeChatModel.builder()
           .apiKey(System.getenv("DASHSCOPE_API_KEY"))
           .modelName("qwen-max")
           .build();
   
       Toolkit toolkit = new Toolkit();
       toolkit.registerTool(new SimpleTools());
   
       ReActAgent agent = ReActAgent.builder()
           .name("Assistant")
           .sysPrompt("You are a helpful AI assistant.")
           .model(model)
           .toolkit(toolkit)
           .build();
   
       Msg userMessage = Msg.builder()
           .role(MsgRole.USER)
           .textContent("Please tell me the current time.")
           .build();
   
       Msg response = agent.call(userMessage).block();
       System.out.println("Agent Response: " + response.getTextContent());
   }

• Installation
• Key Concepts
• Create ReAct Agent

• Model Integration
• Tool System
• Memory Management
• Hook System

• Multi-Agent Pipeline
• State & Session Management
• Multimodal (Vision/Audio)
• Structured Output
• MCP Integration
• RAG
• Planning
• AgentScope Studio

In December, we will further release best practices for context management and reinforcement learning based on Trinity-RFT.

On the technical evolution front, we are continuously exploring more efficient and intelligent context engineering and multi-Agent collaboration paradigms, committed to supporting the construction of more powerful AI applications.

Additionally, addressing the "80/20 rule" characteristic of Agent traffic (where the top 20% of Agents handle 80% of traffic), we will fully advance Serverless architecture, achieving millisecond-level cold starts and hybrid deployment to help developers handle high concurrency while significantly reducing deployment costs and improving efficiency.

AgentScope is released under Apache License 2.0.