Alibaba Launches AgentScope 2.0 to Enhance Agent Deployment Security and Stability

According to Beating Monitor, Alibaba’s Tongyi Lab has released AgentScope 2.0, a multi-agent development framework. Unlike version 1.0, which emphasized transparent development with a focus on visualizing message flows, version 2.0 is centered on high availability and security control for agents in real-world production environments, addressing pain points such as long-task interruptions and unauthorized permission escalation. At the invocation and messaging layers, the framework introduces retry and fallback model mechanisms to prevent task failures due to single-model timeouts. The messaging module has been restructured into Content Blocks that support multimodal streaming data, integrated with an event system to enable streaming output and human-in-the-loop confirmation. To mitigate security risks associated with autonomous agent operation, version 2.0 introduces a new permission system capable of statically and dynamically blocking high-risk shell commands and sensitive file read/write operations. Additionally, version 2.0 introduces a Workspace environment abstraction that decouples agent business logic from execution environments. The same agent can run unchanged across local machines, Docker containers, or E2B cloud sandboxes; combined with a warm-up pool mechanism, this reduces container creation overhead in high-concurrency scenarios such as reinforcement learning. Currently, the Python version of the framework has been upgraded to 2.0, and a brand-new TypeScript version has been launched, with a Java version expected to follow shortly.