// the take
AgentScope 2.0 is a Python framework for building LLM agents, spanning a low-level SDK (ReAct loop, tool/MCP management, memory, middleware) and a full 'agent service' with a FastAPI backend, persistence, and a prebuilt web UI. It's aimed at teams building production agent apps who want serving infrastructure included rather than built from scratch, not hobbyists writing a quick prompt-chaining script.
Spans the whole stack from SDK primitives to a deployable multi-tenant service with session persistence and scheduling, so you're not bolting on your own serving layer. Wide built-in model coverage (OpenAI, Anthropic, Gemini, DashScope, DeepSeek, Moonshot, xAI, Ollama) plus realtime voice over WebRTC and A2A protocol support out of the box. Sandboxed tool execution backends (Docker, E2B, Daytona, K8s, Bubblewrap) are already wired in instead of left as an exercise for the user. Backed by Alibaba with two arXiv papers behind the design and a weekly feature cadence.
Requires Python 3.11+, which rules out plenty of existing deployment environments. The feature surface is large — teams, SOP, scheduling, resource hubs, channels, permission middleware — so there's real ramp-up before you're productive, and the middleware/permission layering adds indirection when something breaks. The agent service drags in a pnpm/TypeScript frontend alongside the Python backend, meaning more moving parts to operate than a plain SDK. Realtime voice, SOP, and model routing are still flagged 'Experimental' in the README, so expect breaking changes in those areas.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →