Player not loading? Watch on YouTube
AgentScope 2.0 is introduced as an open source framework for building AI agent systems. The presentation focuses on execution controls and deployment, using an agent that reads the framework's source code and writes documentation for its permission system.
The first demonstration gives the main model a fake API key. The speaker reports that the task continues through a backup model, illustrating built-in fallback and retry at the model layer. For longer workflows, version 2.0 maintains a structured summary rather than only compacting conversation history.
A unified content block underpins the revised message module. The speaker describes support for different data types and model APIs, while the event system exposes execution steps and supports human approval within the framework.
The permission system considers both the tool and the intent of a command. In the example, read-only operations proceed automatically; writes and uncertain actions require approval. The speaker says the system remembers approvals for later operations.
The workspace demonstration runs the same agent task in Docker through a configuration change. Agent service capabilities move into the core framework, including scheduled cron jobs, and middleware allows custom logic during execution. The presentation does not specify hardware requirements or demonstrate offline model inference. Its claims about security and scalability remain the speaker's claims rather than independently tested results.
