AgentScope is an open-source stack for building, running, and evaluating AI agents across the full lifecycle, from core frameworks and runtimes to memory and training.

Its core framework supports transparent, production-ready agents with tools, planning, voice, and deployment from local environments to Kubernetes, while HiClaw provides a collaborative multi-agent OS for human-in-the-loop task coordination. OpenJudge delivers a unified evaluation framework with production-grade judges that generate training rewards, and additional applications and playgrounds (such as Pawfriends) plus the Spark design system support real-world agent interactions and streamlined developer experience.