agentscope-java
Build distributed, production-grade, long-running agents.
AgentScope Java builds distributed agents
AgentScope Java 2.0 is a production ready Java framework for long running, distributed, enterprise grade agents.
What the framework gives you
AgentScope Java 2.0 is a production ready framework for building distributed, enterprise grade agents. The README describes it as providing the abstractions that work with rising model capability and built in support for long running, safely controlled agent execution. The 2.0 line reached general availability in July 2026, and the README calls it the first production ready release, with a dual layer agent architecture, an event stream, a permission system, middleware, a workspace sandbox, multi agent orchestration, and distributed deployment all ready. The target user is a Java shop that wants agents in existing services rather than a separate Python stack. The framework runs on JDK 17 or newer and is published to Maven Central, so it drops into a standard Java build. The stated aim is enterprise grade behavior: agents that can run for a long time, recover across replicas, and be observed, rather than short scripts that vanish when a process ends. The README points to a documentation site at java.agentscope.io for the building block details and the production guide. The phrase production ready is backed by naming the specific capabilities that make it so, so a reader is not asked to take the label on faith but can look at the event stream, permission system, and distributed deployment as the concrete evidence that the framework is meant for real services.
Core building blocks
The README highlights several building blocks. An event system provides a unified event stream with 31 typed events for real time frontend rendering and human in the loop. A permission system gates tool calls with allow, require user approval, or deny, so an agent's actions can be constrained by policy. Middleware offers aspect oriented hook interception to extend the reasoning acting loop without rewriting the agent. A workspace and sandbox section covers where code and data run. Multi agent orchestration supports several subagent definition patterns with agent_spawn and agent_send and real time event forwarding, so one agent can delegate to others and the events stay visible. Distributed deployment gives true distributed session and memory management across Redis, MySQL, PostgreSQL, OSS, and COS, with cross replica session recovery. The README also mentions an Agent Evolution area with observability, auditing, evaluation, experimentation, and asset management. These pieces together address the parts enterprises worry about most: control, visibility, and recovery when a long running agent spans machines. The 31 typed events are a concrete number that signals the event model is rich rather than a single generic message type, and the allow, require approval, deny triad shows the permission system is built for human oversight rather than full autonomy by default.
From release candidates to GA
The news section shows a steady march from release candidates to the 2.0.0 GA in July 2026. The RC line added model provider modularization, unified multimodal DataBlock support, native structured output, channel integration for IM platforms, subagent cross replica routing, scheduled wakeup dispatching, a fully stateless agent with one line distributed backend, and A2A and AG-UI protocol support. An AgentScope Service component is described as an agent control plane and dashboard for observation, registration, and orchestration, designed to work with AgentScope, LangChain, ADK, and Claude or Qoder. The project also ships a Chinese homepage and documentation alongside the English ones. The README badge states the license as Apache 2.0, and the package is available on Maven Central. For a team adopting it, the RC history is a signal that the APIs were iterated in public before the stable cut, and the GA mark means the building blocks listed above are presented as supported rather than experimental. The breadth of protocol support, including A2A and AG-UI, shows the framework is meant to interoperate with other agent ecosystems rather than lock a team into one vendor's format, which is a meaningful consideration for any enterprise planning a multi year investment in agent infrastructure.
Editorial conclusion
The framework is released under the Apache 2.0 license, written in Java, and published to Maven Central as io.agentscope.
Community notes