Alibaba’s Tongyi Lab has released AgentScope 2.0, an update to its multi-agent development framework that moves beyond the "transparent development" visualization focus of version 1.0. The new version goes all-in on high availability and security for agents running in real production environments — addressing pain points like long-running task interruptions and runaway permissions.
The framework now includes retry and fallback model mechanisms at the call and message layers, so a single model timeout doesn’t bring down the whole task. The message module has been rebuilt as a Content Block that supports multimodal streaming data, paired with an event system for streaming output and human-in-the-loop confirmation. And to keep autonomous agents from doing something stupid (or dangerous), AgentScope 2.0 adds a permission system that can statically and dynamically block high-risk shell commands and sensitive file reads or writes.
There’s also a new Workspace environment abstraction that decouples an agent’s business logic from its execution environment. The same agent can run locally, in a Docker container, or in an E2B cloud sandbox without modifying any code. A pre-warming pool mechanism reduces container creation overhead for high-concurrency scenarios like reinforcement learning. The Python version has already been upgraded to 2.0, a TypeScript version is out now, and a Java version is coming soon.