AgentScope Java 1.0 Empowers Java Developers to Build Enterprise‑Grade Agentic Apps

ReAct development paradigm

AgentScope adopts the ReAct (Reason‑Act) paradigm. The LLM acts as the reasoning and planning brain and decides autonomously when and how to invoke tools, providing flexibility for complex tasks compared with a fixed workflow where the tool‑collaboration path must be pre‑defined.

Real‑time intervention

An asynchronous architecture enables safe interruption, context‑preserving pause, and immediate termination when a task deviates or exceeds time limits. Developers can plug custom interruption‑handling logic for fine‑grained control.

Efficient tool calling

Standardised registration interface extracts a tool’s JSON‑Schema, supports parameter presets and post‑processing.

Tools are organised into functional groups (e.g., browser, map) and meta‑tools that allow dynamic activation, reducing context‑window pressure.

Unified asynchronous streaming interface handles sync, async or streaming outputs uniformly and supports parallel tool calls for higher throughput.

Built‑in tools

PlanNotebook provides task planning, creation, modification, pause, resume and switching of multiple plans, enabling orderly execution of complex workflows.

Structured output tool enforces JSON‑format responses from LLMs, eliminating manual prompt engineering and secondary parsing.

Enterprise‑grade capabilities

Security sandbox isolates tool execution. Three sandbox types are provided: GUI sandbox with full desktop interaction, filesystem sandbox with isolated file read/write, and mobile sandbox based on an Android emulator supporting touch, swipe, input and screenshot.

Context engineering includes:

RAG: built‑in embedding‑based retrieval, optional private‑cloud knowledge‑base deployment, and integration with Alibaba Cloud Baichuan knowledge base.

Memory: abstractions for short‑term and long‑term memory, semantic search, multi‑tenant isolation, and three control modes (automatic management, agent‑initiated calls, hybrid).

Integration is simplified through MCP (HTTP service) and A2A (Agent Card registration via Nacos), allowing agents to be called like micro‑services or to discover other agents automatically.

Performance optimisations

Lightweight core depends only on Reactor Core, Jackson and SLF4J; RAG and memory are optional extensions.

Asynchronous message queue (RocketMQ) decouples tasks and boosts throughput.

Native optimisation with GraalVM and Leyden achieves 3‑10× faster cold start (≈200 ms), enabling millisecond‑scale serverless AI workloads.

AI‑native application architecture

Non‑deterministic agents replace deterministic software, making effect evaluation data‑driven. A/B testing becomes essential for version quality. AgentScope supplies Studio, RM Gallery and Trinity‑RFT together with the Higress AI gateway to close the data‑flywheel: collect live data, evaluate with reward models, and continuously fine‑tune models.

Roadmap

Future work includes continuous context‑engine optimisation, real‑time multimodal support, and lowering the barrier for evaluation and reinforcement‑learning pipelines.

GitHub repository: https://github.com/agentscope-ai/agentscope-java

Documentation: https://java.agentscope.io/en/intro.html