agentscope-ai/agentscope

What it solves

AgentScope is a production-ready framework designed to build agentic applications. It moves away from strict prompt-based orchestration, instead leveraging the inherent reasoning and tool-use capabilities of modern LLMs to create flexible, autonomous agents.

How it works

The framework is split into two primary layers:

1. SDK Layer: Provides building blocks for composing agents, including a ReAct reasoning-acting loop, toolkits for Python tools and MCP servers, support for multiple LLM providers, context management (compaction and injection), and isolated execution environments (sandboxes) like Docker or E2B.
2. Agent Service: A FastAPI-based backend that transforms agents into multi-tenant, multi-session applications. It includes a pre-built Web UI, RAG services, and the ability to connect agents to IM platforms like Discord and Feishu.

Who it’s for

Developers building AI agents and multi-agent systems that require production-grade features like multi-tenancy, session isolation, persistence, and secure code execution.

Highlights

• Comprehensive Tooling: Built-in support for shell, file editing, and search tools, as well as integration with MCP servers.
• Agent Teams: Support for leader-worker orchestration where a leader agent delegates tasks to members.
• RAG Integration: Built-in RAG service with blob storage and multi-tenant retrieval.
• Realtime Capabilities: Experimental support for realtime voice conversations over WebRTC and speech-to-speech APIs.
• Secure Execution: Multiple sandbox options for isolated tool and code execution, including K8s and Docker.