agentscope-ai/agentscope-java

AgentScope Java 2.0 is worth a serious look if you are building agents in Java and need them to survive beyond a single request. Teams not already in the JVM ecosystem, or prototyping small single-turn assistants, can reasonably skip it.

The problem

A team wants an assistant that keeps working across sessions: it should remember earlier decisions, pause for human approval before a risky tool call, and stream progress into an existing product UI. Building this on a bare ReAct loop means writing conversation persistence, permission checks, event plumbing and compaction by hand, then again for each new agent. AgentScope packages these into one harness: you configure a shared builder at startup and create an agent per request, with identity and storage carrying the conversation across instances. The result is an agent that can be resumed and observed rather than a one-shot script.

What's different

The framework layers engineering infrastructure on top of the ReAct reasoning core through middleware and toolkit channels, leaving the reasoning loop untouched. Two pieces stand out in the README: a 32-event typed stream for real-time rendering and human-in-the-loop work, and a permission system that gates each tool call as allow, require approval or deny. Persistent memory is split between context, an agent-curated MEMORY.md, and an on-disk fact log, with automatic compaction to bound prompt size. There is also a self-evolution mechanism where successful patterns are saved as Markdown skills and loaded on demand.

The trade-off is that this is a full platform rather than a thin library. Adopting it means accepting its abstractions, its file-based state layout and its deployment model, so evaluate whether your agent's complexity justifies that weight. Model providers are separate extension modules, so the provider you use determines an extra dependency and possibly provider-specific behavior.

Who should try it

Good fit: a backend team with JDK 17+ services that needs agents tied into an existing application through endpoints, where background tasks, interactive approvals and reviewable delivery are part of the workflow. The separate AgentScope Service and control plane are aimed at exactly that: publishing capabilities through APIs while the application keeps its own interface and authorization.

Poor fit: a developer who wants a minimal conversational agent in Python or a lightweight script, or who only needs one reasoning turn. The framework's distributed, long-running orientation adds concepts those users will not use. A team committed to another agent stack may still find the event and permission designs useful to read, but should not expect a drop-in replacement.

Requirements and limitations

JDK 17 or higher is required. Model providers ship as separate Maven modules; the quickstart uses DashScope and expects an API key, so the initial setup includes choosing and wiring a provider rather than just adding one dependency. The main artifact is agentscope-harness at version 2.0.3, though checked-in examples may trail the release. Ongoing upkeep means managing persisted agent state, skill files and memory files alongside your application data, plus reviewing which tool calls are permitted.

The README is excerpted and omits details such as the full module list and production operations guidance. It does not document clustering, state-store backends or upgrade paths for persisted state, so a production team should confirm those against the documentation before committing. v2.0.0 is the first production-ready release, per the release notes, though that is the project's own framing.

HotRepo's take

HotRepo records the project at 6,136 stars on 2026-10-27, with +12 that day and a single day on the daily trending board. That is one dated snapshot, so it neither shows a trend nor supports claims about persistence. Treat the framework on its design and documentation: start with the permission system and event model docs to check whether your human-in-the-loop requirements fit. Before adopting, verify the documented state-storage options for long-running agents.