AgentTeams
About
Multi-agent setups usually hide agent-to-agent traffic in black-box calls, and leaked API keys from a compromised worker are the standard failure mode. AgentTeams, from the AgentScope team, instead places a Manager agent and its Workers in self-hosted Matrix rooms alongside the human, with declarative Worker/Team/Human definitions and runtime swaps (OpenClaw, QwenPaw, Hermes) via commands like agt update worker --runtime hermes. Security rests on the Higress AI gateway holding every real credential — workers receive consumer tokens only, so a compromised worker exposes no LLM, GitHub, or MCP keys. A MinIO-backed shared filesystem reduces token spend between agents. Deployment is a one-command install script or a Helm chart bundling Higress, Tuwunel, MinIO, and the controller, with the project backed by Alibaba/Aliyun.
Other entries by agentscope-ai
- AgentScope — agent-sdk