Community

Blog Events Webinars Tutorials Forum
  1. Blog
  2. Events
  3. Webinars
  4. Tutorials
  5. Forum
Create Account
×
Community Blog Agentscope Java v1.0 Has Been Released, Allowing Java Developers to Easily Build Enterprise-Level Agentic Applications

Agentscope Java v1.0 Has Been Released, Allowing Java Developers to Easily Build Enterprise-Level Agentic Applications

This article introduces AgentScope Java 1.0, a framework that enables Java developers to easily build enterprise-grade agentic applications.

By Yi Zhan

Introduction to AgentScope

AgentScope is an open-source framework launched by Alibaba that is developer-centric and focuses on intelligent agent development. It is a strategic product at the Agent layer following ModelScope (Magic Community) . Its core goal is to solve the challenges faced by intelligent agents in construction, operation, and management, providing a production-grade solution that covers the full lifecycle of "development, deployment, and optimization," making the development of intelligent agent applications simpler, more stable, and more effective.

1

Recently, AgentScope has welcomed a major update for the December version, which is a significant infrastructure upgrade aimed at production-grade intelligent agent applications, allowing agents to transition from "lab prototypes" to "business implementation." This update revolves around three core themes: ready-to-use intelligent agents that instantly empower various real-world scenarios; infrastructure enhancements that comprehensively upgrade the underlying capabilities that make agents "smarter"; runtime × multi-language × front-end, delivering production-ready agents in a three-in-one manner.

For a long time, the Java language has dominated development in the fields of finance, government affairs, and e-commerce. The developer community has also strongly called for a Java version of AgentScope, and now it has significantly released the 1.0 version of Java, embracing the mainstream technology stack for enterprise development.

AgentScope Java 1.0 Major Release

Today, we are pleased to announce that the AgentScope Java 1.0 version is officially released, providing enterprise-grade capabilities for building Agentic applications for Java developers.

2

Firstly, in terms of development paradigm, AgentScope adopts the leading ReAct (Reasoning-Action) model, supporting efficient tool invocation and allowing developers to intervene in the agent execution process in real time, achieving a perfect balance between autonomy and controllability.

Secondly, it provides ready-to-use enterprise-grade capabilities. The framework offers a secure sandbox to guarantee the safety of code execution, optimizing model interaction effects through fine-tuned contextual engineering. As a Java framework, it is easy to integrate into existing enterprise technology stacks and boasts a high-performance architecture, ensuring stability and reliability in production environments.

Finally, it possesses a complete ecosystem for development and optimization. It offers a complete toolchain from visual debugging in development mode, A/B testing, to evaluation and reinforcement learning, forming a closed loop for agent development, deployment, and optimization, helping to continuously improve agent effectiveness.

Leading Development Paradigm

When building complex AI agent applications, developers generally face numerous challenges: rigid workflows that are difficult to adapt to changing tasks, real-time intervention in running agents, chaos in management due to a plethora of tools, unstable model output formats, etc. Systematically solving these pain points is key to enhancing development efficiency and application stability. AgentScope adopts the leading ReAct paradigm, granting LLM autonomous planning capabilities and providing real-time intervention controls and an efficient tool invocation system. Additionally, it integrates powerful tools such as task planning and structured outputs, supporting efficient development of production-grade applications.

1.  The leading ReAct paradigm empowers agents with autonomous planning capabilities.

  • Workflow mode: In this mode, the collaboration path between LLM and tools is predefined by developers, who have complete control over the system's execution flow, ensuring task execution stability and determinism. However, its drawback is a rigid architecture that incurs high maintenance costs when business logic becomes complex and prevents enjoying the capabilities provided by LLM's continuous evolution.
  • ReAct paradigm: In contrast to Workflow, ReAct grants LLM autonomous control. The LLM acts as the brain, dynamically reasoning and planning, autonomously deciding when and how to invoke tools to perform actions, thus dominating task completion. As LLM matures in key abilities like understanding, planning, and tool usage, this highly autonomous agent architecture has become the preferred choice in complex application scenarios.

2.  Real-time intervention control allows for full control over agent operation. Traditional agents become uninterruptible once started, but AgentScope uses an asynchronous architecture to implement a powerful real-time intervention mechanism.

  • Safe interruption: Supports pausing the agent at any time and automatically saving its context and tool status to ensure tasks can resume seamlessly.
  • Real-time interruption: When tasks deviate from expectations or exceed time limits, users can terminate them immediately to avoid resource waste.
  • Flexible customization: Developers can define custom interruption handling logic, achieving more refined management.

3.  Efficient tool invocation: With the explosion in the number of callable tools, agents face issues of complex tool management, low execution efficiency, and tense context. AgentScope has built a reliable and efficient tool management system.

  • Tool registration: Provides a standardized registration interface, supports automatic extraction of tool JSON Schema, offers parameter presets, and tool function post-processing interfaces, lowering integration barriers.
  • Convenient management: AgentScope efficiently supports tool usage through structured organization methods and dynamic control mechanisms. Tool groups classify tools based on functionality (e.g., browser, mapping service), enabling agents to activate related tools as needed for current tasks, thus alleviating context window pressures; meta-tools allow agents to dynamically enable or disable entire tool groups during runtime, achieving smarter tool management.
  • Efficient execution: Adopts a unified interface for handling all tool calls, whether synchronous, asynchronous, or streaming output, which will be unified as asynchronous streaming returns in AgentScope, reducing the complexity of processing tool function returns while supporting parallel tool calls to significantly enhance operational efficiency.

4.  Powerful built-in tools: AgentScope includes many powerful ready-to-use tools that accelerate the development of production-grade applications.

  • The PlanNoteBook tool provides robust task planning and execution capabilities. It allows developers to manually define structured plans, and also permits agents to autonomously create and manage plans during runtime. PlanNotebook offers complete plan management functionality, including creating, modifying, pausing, resuming, and switching multiple plans, guiding agents to execute complex plans in an orderly manner.
  • Structured output: Traditional approaches require specifying format requirements in a Prompt, asking the model to "please output in the following JSON format," and making repeated trials and optimizations of prompts, often needing secondary parsing and format validation in external code. AgentScope ensures that LLM outputs strictly follow predefined JSON formats through built-in tools, completely eliminating cumbersome prompt debugging and secondary parsing.

Enterprise-Grade Capabilities

AgentScope provides a secure tool sandbox and contextual engineering capabilities, addressing the core pain points of safety and effectiveness, ensuring that the agent's output meets production standards. Leveraging Java's robust ecosystem in enterprise application development, it offers flexible integration and embedding solutions through standard A2A and MCP protocols, allowing agents to be embedded as independent services in existing systems or to serve as intelligent hubs for connecting and orchestrating other services. Developers need not focus on underlying integration details but can quickly construct production-grade agent applications by concentrating on business logic. Lastly, relying on the capabilities provided by AgentScope Runtime, it supports one-click deployment of agents to Alibaba Cloud Hundred Refinement and function computing platforms, offering commercial-level productization assurance for your agent applications.

1.  Secure sandbox:

  • When agents perform tool calls or automation tasks, they may access sensitive resources or trigger uncontrollable behaviors, necessitating a sandbox that provides a secure isolated environment. AgentScope Runtime Sandbox allows developers to deploy custom tools to run safely in a highly isolated controlled environment, preventing unintended effects or security risks on the system.
  • Built-in various ready-to-use sandboxes: The GUI sandbox provides a complete desktop environment, supporting mouse, keyboard, and screen operations; the file system sandbox achieves isolated file read and write management; the mobile sandbox based on Android emulators supports real mobile interactions such as clicks, swipes, inputs, and screenshots. It balances security, flexibility, and multi-platform coverage, fully supporting complex scenarios like tool execution, browser automation, and training evaluation.

2.  Contextual engineering:

  • RAG: Built-in standard implementation based on embedding supports enterprises in facing complex, diverse business data scenarios to privately deploy their knowledge base systems, achieving complete control over data; integrating Alibaba Cloud Hundred Refinement’s enterprise-level knowledge base, leveraging commercial products to obtain stronger retrieval and reordering capabilities.
  • Memory: AgentScope defines abstractions for short-term and long-term memory, supporting semantic search and multi-tenant isolation, providing three control methods—automatic management, proactive calls by agents, and mixed mode. The ReMe project offers optimal practices for memory, enabling agents to understand user preferences, enhance task performance, and utilize tools more intelligently, significantly improving the accuracy and contextual coherence of intelligent Q&A in business scenarios, achieving better performance over time.

3.  Easy Integration

  • MCP integration: Based on the AgentScope Java open-source ecosystem, existing HTTP business systems do not require changes to business logic code and can be seamlessly integrated with agents through simple configuration, rapidly becoming "hands and feet" callable by agents, greatly expanding the agent's capability boundaries.
  • A2A integration: Complex tasks often require multiple agents to work together. AgentScope Java supports registering agent cards that describe the capabilities of agents to service centers like Nacos. The calling agent only needs to connect to Nacos to automatically discover and invoke other agent capabilities. This makes building and collaborating complex distributed multi-agent systems as simple as calling ordinary microservices.

4.  High Performance

  • Lightweight: The core library relies only on Reactor Core, Jackson, and SLF4J, while RAG and long-term memory capabilities are introduced as optional extensions. Currently, it implements model calls based on the vendor's native SDK, with plans to simplify kernel dependencies further by moving to OkHttp and Jackson's native implementations in the future.
  • Asynchronous: In response to the characteristics of AI application interactions, which involve long durations and multiple rounds of contextual dependencies, it supports introducing a message queue RocketMQ as an asynchronous message hub, decoupling tasks and enabling non-blocking calls, enhancing the agent's throughput capability and response speed.
  • Native optimization: Together with the JVM team, we have adapted GraalVM and Leyden to improve the startup speed of Java applications by 3 to 10 times, achieving cold starts of agents within 200ms, laying the foundation for millisecond-level elasticity in serverless AI applications.

Robust Ecosystem

The AI-native application architecture is profoundly reshaping software engineering paradigms, replacing the determinism of traditional software with the non-determinism of agents, where the final effect is jointly determined by models, data, and context. This evolution transforms traditional "code testing" into complex "effect evaluation." Since any minor change can lead to drastic fluctuations in effect, A/B testing has transitioned from an optimization option to a core release process for ensuring version quality. Therefore, the focus of software engineering must shift from code-centric to data-centric, with success hinging on building an efficient data flywheel.

In the face of this challenge, AgentScope provides a series of ecological tools such as Studio, RM Gallery, and Trinity-RFT, combined with the Higress AI gateway and observable systems, enabling you to quickly practice the AI-native application data flywheel.

3

During the agent development phase, we use the AgentScope Studio visualization platform for real-time debugging and observation of agents, significantly enhancing development efficiency. It integrates deeply with OpenTelemetry and LoongSuite, achieving end-to-end full-link tracking.

In the deployment architecture, Higress serves as a unified traffic entry gateway, responsible for routing external requests to the corresponding agents. Agents communicate with LLM through Higress's built-in AI gateway capabilities. With Higress's powerful plugin system, we can flexibly label traffic to achieve precise routing control over agents and LLMs.

During the post-release A/B testing phase, Higress gateway can allocate traffic to different experimental groups based on request content (such as user geographic location, business line, and payment status). For example, paying users could be directed to version A of the agent, while free users are directed to version B for effect comparison. Additionally, the grouping labels of traffic, with observability, will pass through the entire call chain. Thus, the AI gateway can route requests to the corresponding LLM versions based on these labels. This mechanism allows us to perform coordinated A/B testing of agents and LLMs without modifying business code.

Throughout this process, all data generated across the full link—from user input, agent prompts, to model output, latency, and costs—will be reported to the observable system. Based on the RM Gallery's reward function, agent performance across experimental groups will be evaluated and high-quality datasets will be filtered out. Subsequently, our training framework, Trinity-RFT, will use these datasets and reward models to continuously iterate and enhance the model's ability to solve business problems.

Ultimately, this forms a self-optimizing loop driven by data. The system continuously collects real online data, analyzes and evaluates effects, and transforms that into high-quality training data, thereby continuously strengthening model capabilities and building a solid technical competitive barrier.

AgentScope Java Roadmap

  • Continuous optimization of contextual engineering: The reasons for insufficient agent effectiveness are either due to weak model capabilities or inaccurate contextual provision. Contextual engineering is at the core of engineering capabilities. AgentScope Java will continue to delve into contextual engineering, aiming to build a more efficient and low-latency contextual management system. In the future, developers will not need to concern themselves with the technical details of context, only needing to focus on defining the functionality of agents.
  • Real-time multimodal support: The boundaries of large models are expanding from text to images, speech, and even video. Interactive intelligent products that can engage with the physical world are beginning to enter our lives. AgentScope Java will build deep support for real-time multimodal applications, helping developers better create multimodal applications, where in the future, agents won't just take text input, but can also serve users better through "eyes," "ears," and "hands."
  • Evaluation and reinforcement learning optimization: We have already provided a comprehensive solution for observation, evaluation, and optimization, but the current threshold for evaluation and reinforcement learning remains relatively high. We will continuously lower the threshold through ecological integration, allowing developers to write business logic and design reward functions, using AgentScope's ecological toolchain to enable agents to evolve through interactions with users or environments, achieving real self-growth and iteration.

We welcome you to join our open-source community to build an agent system for the future together!

java react AI Artificial Intelligence Developers Open Source Web Development cloud native Generative AI GenAI LLMs AI Agent AgentScope Reasoning-Action
0 0 0
Share on

Read previous post:

Explore Mobile Performance Monitoring: Architecture and Implementation of an Observability Agent for Android

Alibaba Cloud Native Community

629 posts | 55 followers

You may also like

Comments

Alibaba Cloud Native Community

629 posts | 55 followers

Related Products

More Posts by Alibaba Cloud Native Community

See All