AgentScope is an open-source agent framework with 33,500+ stars on GitHub for building LLM agents you can see, understand and trust.
AgentScope, from Alibaba’s Tongyi team, is a full platform for building, running and observing multi-agent systems in Python. Its pitch is transparency: agents are composed from explicit building blocks, their conversations and tool calls are visible, and a visual studio helps you debug behavior instead of guessing from logs. The framework covers single agents, multi-agent pipelines (including leader-delegates-to-members TeamPipeline), realtime voice agents over WebRTC, and integrations via MCP. It is model-agnostic, with first-class support for DashScope, OpenAI, Gemini, MiniMax and other providers, plus middleware such as model routing. Development moves fast — realtime voice and team pipelines landed in autumn 2026 — and the repository was updated in October 2026.
The framework reflects hard lessons from production agent deployments: failures should be observable, conversations replayable, and components replaceable when a better model appears. Its building blocks cover the full lifecycle — prompt formatting, tool calling, memory, multi-agent messaging and evaluation — so teams are not forced to bolt five libraries together. English and Chinese documentation are both first-class, and the project’s release cadence, with meaningful features landing monthly, has made it one of the most-watched agent frameworks to come out of a large lab’s open-source program.
AgentScope is free and open source under the Apache License 2.0. You pay only for the model APIs you choose to use.
Python developers and research teams building multi-agent applications who want observability and control rather than a opaque agent framework.