Agentscope Java
AgentScope Java: Agent-Oriented Programming for Building LLM Applications
Last verified:
What is Agentscope Java?
AgentScope Java is an agent-oriented programming framework for the JVM that brings reasoning, tool integration, memory management, and multi-agent collaboration to Java enterprise applications. It is the Java port of AgentScope, aligned with the Python version in design and features, and serves as the underlying engine for Spring AI Alibaba's Agentic APIs.
Key features include out-of-the-box ReAct agents with reasoning and acting capabilities, comprehensive tool calling with sync/async support and parallel tool calls, MCP server integration for service connectivity, modular memory management with both short-term conversation history and long-term persistent memory (ReMe), multi-agent collaboration patterns including "Agent as Tool" for hierarchical systems, multimodal capabilities supporting images/audio/video via ContentBlock architecture, RAG integration with multiple backends (Bailian, Qdrant, Dify, RAGFlow, HayStack), session storage options (MySQL, Redis), planning capabilities via PlanNotebook for breaking down complex tasks, streaming responses for real-time visibility, AG-UI protocol integration for frontend-backend communication, and Spring Boot/Quarkus/Micronaut framework support.
AgentScope Java is designed for Java developers building LLM-poweredagent applications, enterprise teams needing multi-agent systems with tool integration, developers migrating from Python AgentScope to JVM, Spring Boot developers seeking agentic APIs through Spring AI Alibaba, and anyone requiring production-ready agent frameworks with comprehensive documentation and community examples.
Agentscope Java pricing
Pricing model: Freemium
AgentScope Java is open-source and free to use. The framework is available via Maven Central withGroupId io.agentscope. Version 1.0.5 is the current release. No paid tiers or licensing fees are mentioned. The all-in-one package includes DashScope SDK and MCP SDK by default at no additional cost beyond API keys for LLM providers. Users only pay for their chosen LLM API usage (DashScope/Qwen, OpenAI, Anthropic) and any third-party services like Qdrant, Redis, or MySQL for RAG and session storage.
Agentscope Java pros
- Open-source multi-agent platform with active development
- ReActAgent provides out-of-the-box reasoning and tool calling
- Supports multiple LLM providers (DashScope/Qwen, OpenAI, Anthropic)
- MCP SDK included for Model Context Protocol integration
- Modular extension system with core + extensions dependency model
- Long-term memory via ReMe and Mem0 extensions
- Multiple RAG backend options (Qdrant, Dify, RAGFlow, HayStack)
- Session persistence with MySQL and Redis support
- Multimodal support for images, audio, and video
- Agent as Tool pattern for hierarchical multi-agent systems
- PlanNotebook for automated task planning and decomposition
- Streaming responses with real-time event visibility
- AG-UI protocol integration for web frontend connectivity
- Spring Boot starter with auto-configuration support
- Quarkus and Micronaut extensions available
- JSON Session for persistent state storage to files
- Tool groups for organizing expert agents
- Structured output support with JSON Schema generation
- User interruption and custom interrupt handling
- Parallel tool calls for improved efficiency
Agentscope Java cons
- Requires JDK 17 or higher
- Version 1.0.5 is relatively new with potential API changes
- Agent as Tool is marked as experimental feature
- All-in-one package may bloat projects with unnecessary dependencies
- DashScopeChatFormatter required for Qwen vision models
- Drag-and-drop Workstation UI no longer maintained since 1.0
- Not source-compatible with AgentScope 2.0 Python version
- Base64 encoding recommended for best media compatibility adds overhead
- InMemorySession loses state on process restart by default
- Learning curve for agent-oriented programming paradigm
Frequently asked questions about Agentscope Java
What is AgentScope Java?
AgentScope Java is an agent-oriented programming framework for the JVM that brings reasoning, tool integration, memory management, and multi-agent collaboration to Java enterprise applications. It is the Java port of AgentScope, aligned with the Python version in design and features.
What are the system requirements for AgentScope Java?
AgentScope Java requires JDK 17 or higher. The all-in-one package includes DashScope SDK, MCP SDK, Reactor Core, Jackson, and SLF4J by default. Additional dependencies may be needed for other models or features like OpenAI, Qdrant, or PDF processing.
How do I install AgentScope Java?
You can install via Maven with dependency io.agentscope:agentscope:1.0.5 (all-in-one) or io.agentscope:agentscope-core:1.0.5 (core + extensions). Gradle users can use implementation 'io.agentscope:agentscope:1.0.5'. The all-in-one package is recommended for most users.
What LLM models does AgentScope Java support?
AgentScope Java supports multiple LLM providers including DashScope (Qwen series like qwen-plus, qwen-vl-max, qwen-audio-turbo), OpenAI (gpt-4o, gpt-4-vision-preview), and Anthropic (claude-3-opus, claude-3-sonnet). Each model requires its corresponding SDK dependency.
What is the ReActAgent?
ReActAgent is an out-of-the-box ReAct (Reason + Act) agent provided by AgentScope Java. It supports hooks around reasoning and acting, structured output, user interruption, sync/async tool functions, streaming tool responses, parallel tool calls, MCP server integration, short-term memory, and long-term memory with AGENT_CONTROL, STATIC_CONTROL, or BOTH modes.
How does multi-agent collaboration work?
AgentScope Java supports multi-agent collaboration through the 'Agent as Tool' pattern, which allows registering an agent as a tool that can be called by other agents. This enables expert specialization, task delegation, and multi-turn conversations with session state persistence. Sub-agents can maintain conversation state via session_id parameter.
What memory management options are available?
AgentScope Java offers short-term memory for conversation history (InMemoryMemory) and long-term memory via extensions including ReMe (modular memory management with file-based and vector-based storage), Mem0 Long-term Memory, and AutoContext Memory. Memory mode can be AGENT_CONTROL, STATIC_CONTROL, or BOTH.
How does multimodal support work?
Multimodal capabilities enable agents to understand and generate images, audio, video, and other media via the ContentBlock system. Supported formats include images (PNG, JPEG, GIF, WebP), audio (MP3, WAV, MPEG), and video (MP4, MPEG). Media can be loaded via Base64 encoding (recommended) or URL references.
What RAG backends are supported?
AgentScope Java supports multiple RAG backends through extension modules: Bailian RAG, Qdrant RAG, Dify RAG, RAGFlow RAG, and HayStack RAG. Each requires its corresponding extension module which automatically includes required third-party dependencies.
How does AgentScope Java relate to Spring AI Alibaba?
AgentScope Java is the Java port of AgentScope under active development, aligned with the Python version in design and features. Spring AI Alibaba will adopt AgentScope Java as its underlying engine. If you use Spring AI Alibaba's Agentic APIs, you automatically gain AgentScope capabilities after upgrading without needing to integrate AgentScope Java separately.