AgentScope

AgentScope is a production-ready, open-source agent framework built for developers who want to build, understand, and trust their AI agents. It provides essential abstractions that scale with rising model capabilities, supporting everything from simple ReAct agents to complex multi-agent workflows with realtime voice, memory, and reinforcement learning fine-tuning. Licensed under Apache 2.0, AgentScope is installable via PyPI and requires Python 3.10+.

• Built-in ReAct Agent — get started in 5 minutes with a pre-built ReAct agent, tools, and human-in-the-loop steering out of the box.
• MCP & A2A Support — use MCP tools as local callable functions and compose toolkits; supports Agent-to-Agent (A2A) protocol for inter-agent communication.
• Realtime Voice Agent — build voice-enabled agents with web interfaces supporting real-time speech input and output, including multi-agent voice workflows.
• Flexible Memory Module — in-memory and database-backed memory with compression, long-term memory via ReMe, and SQLite session support.
• Multi-Agent Orchestration — MsgHub and pipeline primitives enable sequential, concurrent, and dynamic multi-agent conversations with flexible message routing.
• Agentic Reinforcement Learning — seamlessly fine-tune agents using the Trinity-RFT library with RL training across math, navigation, tool-use, and game scenarios.
• Human-in-the-Loop Steering — supports real-time interruption and resumption of ReActAgent conversations with robust memory preservation.
• Observability & Deployment — deploy locally, as serverless cloud functions, or on Kubernetes with built-in OpenTelemetry (OTel) support.
• Extensive Ecosystem Integrations — integrates with DashScope, Anthropic Agent Skills, RAG pipelines, structured output, TTS, and more.
• Evaluation & Benchmarking — includes ACEBench evaluation support and multiple sample projects for tuning and testing agent performance.