# AgentScope

A production-ready agent framework built around a ReAct reasoning-and-acting loop, a pluggable toolkit and model layer, and composable middleware, designed to work with increasingly capable models rather than constrain them with rigid orchestration.

AgentScope composes an agent from separate building blocks: a ReAct loop with structured output, real-time interruption and resumption, and batched sequential or concurrent tool calling; a toolkit that manages Python tools, MCP servers and skills, including built-in coding tools such as shell, file editing and search; a model layer spanning LLM, embedding and text-to-speech across providers including OpenAI, Anthropic, Gemini, DashScope, DeepSeek and Ollama; automatic context compaction and injection of system prompts, retrieval-augmented generation and memory; an event bus that streams reasoning, tool calls and multimodal content to a frontend; fine-grained permission and human-in-the-loop controls; and switchable memory backends. Isolated execution is available through several sandboxes, including Docker, Apple Container, Bubblewrap, E2B, Daytona and Kubernetes.

Recent additions include model routing that picks a chat model per reply, a realtime voice agent, support for the Agent2Agent (A2A) protocol to talk to remote agents, a terminal console for testing agents, and channel integrations for DingTalk, Feishu and Discord.

It requires Python 3.11 or higher and is licensed under Apache-2.0.

Website: https://docs.agentscope.io/