AgentScope Java 1.1.0 Implements Harness Framework to Scale Enterprise Agents

AgentScope Java 1.1.0 has been released as a milestone implementation of a planned “Harness Framework” for agent development, bundling persistence, isolation and recoverability features intended to make agent logic portable from local tools to distributed systems. The release emphasizes a single runtime model so teams can run the same agent implementations for local productivity agents such as XxxClaw and Coding Agent and for enterprise — grade, multi — replica deployments like DataAgent and SRE Agent. This unified approach matters because it targets the operational gaps that typically block moving prototypes into production.

The maintainers position the release against recent interest in agent architectures and Harness Engineering popularized by projects such as OpenClaw, Hermes and Claude Code. They say teams moving those designs from single — machine prototypes to multi — user, replicated services regularly hit scaling and security barriers. By addressing persistence, isolation and recoverability explicitly, AgentScope aims to remove those practical blockers and to support continuous evolution of agent capabilities while meeting enterprise needs for multi — replica access and tenant isolation.

The runtime centers on a workspace — driven model that consolidates an agent’s persona, memory, skills and sub-agent definitions into a structured workspace. Context is automatically loaded from the workspace before each run and memory is written back afterward, which the project describes as enabling continuous state and capability evolution instead of ephemeral, per-session behavior. The release is presented as a platform for teams adopting Harness Engineering practices in Java.

Storage is handled via a pluggable abstract file system that exposes a uniform interface for switching physical backends. Workspaces can be stored on local disk, remote shared storage, or isolated sandbox volumes without changing higher — level agent logic. Complementing storage flexibility, the release introduces conversation compaction, two-tier memory consolidation, and full-text search to improve context retrieval and reduce conversation bloat. A background maintenance mechanism is included to limit uncontrolled memory growth over time.

Sub-agent orchestration supports declarative definitions and both synchronous and asynchronous delegation of tasks. Tools and skill scripts can be configured to execute inside isolated sandboxes, with sandbox state kept recoverable across multi — turn conversations. That execution model is intended to balance session — level and user-level isolation for multi — tenant deployments while preventing direct execution of arbitrary user commands on the host.