Alibaba Releases AgentScope 2.0: Addressing Environment Decoupling and Security Perimeter Issues During Agent Deployment

According to Dynamic Insight monitoring, Alibaba's Tongyi Lab has released the AgentScope 2.0 version of its Multi-Agent development framework. Contrasting with the 1.0 version's focus on visualizing message interactions through "transparent development," the new version emphasizes the high availability and security control of agents in real-world production environments, aiming to address pain points such as long task interruptions and permission sprawl.

At the invocation and messaging layer, the framework has introduced retry and fallback models to prevent task interruptions caused by single-model timeouts. The messaging module has been refactored to support Content Blocks for multimodal streaming data, along with an event system that provides streaming output and human-in-the-loop capabilities. To mitigate security risks associated with autonomous agent execution, version 2.0 has added a permission system that can statically and dynamically intercept high-risk shell commands and sensitive file read/write operations.

Furthermore, version 2.0 has introduced Workspace environment abstraction, decoupling an agent's business logic from its execution environment. The same agent can run in a local environment, Docker container, or E2B cloud sandbox without code modifications, leveraging a warm-up pool mechanism to reduce container creation overhead in high-concurrency scenarios such as reinforcement learning. Currently, the framework has completed the Python version's 2.0 upgrade and has newly launched a TypeScript version, with a Java version set to follow up with updates soon.