AgentScope Runtime: the spin-off that got merged back in

What it does

AgentScope Runtime was a Python infrastructure layer for turning agent applications into deployed services. It provided AgentApp, a class built by directly inheriting from FastAPI, for streaming agent responses over SSE, and it ran tool calls inside hardened sandboxes covering Python, shell, browser, GUI, filesystem, and mobile contexts. The project also aimed to abstract away deployment targets—local, Kubernetes, or serverless—and offered adapters for several mainstream agent frameworks.

The interesting bit

The runtime tried to solve the “ops” problem for autonomous agents: how do you let a large language model invoke arbitrary tools without trashing the host system? Its answer was a battery of async sandbox implementations and a distributed interrupt service for preempting runaway tasks. Ultimately, the maintainers decided these features belonged in the main AgentScope framework rather than a separate repository, so this standalone runtime was absorbed into AgentScope 2.0 and archived.

Key highlights

• Sandboxed execution for Python, shell, browser, GUI, filesystem, and mobile tool calls.
• AgentApp inherits directly from FastAPI, aiming for native integration with the Python web ecosystem.
• Deployment abstraction for local, Kubernetes, and serverless environments.
• Adapters for AgentScope, LangGraph, Microsoft Agent Framework, Agno, and AutoGen, though support levels vary significantly.
• Full-stack observability hooks for logs and traces.

Caveats

• The repository is archived and read-only; all capabilities now live in AgentScope 2.0.
• Framework compatibility was uneven at archival time: LangGraph and AutoGen integrations were marked as incomplete.
• No further updates or community support will arrive here.

Verdict

Curious developers studying agent sandboxing patterns or maintaining legacy AgentScope Runtime services might find the code instructive. Anyone starting fresh should ignore this repository and use AgentScope 2.0 directly.