agentscope-ai/AgentTeams

What it solves

AgentTeams provides a controlled, auditable environment for multi-agent collaboration. It solves the problem of managing multiple AI agents in enterprise settings where security, human oversight, and visibility are critical, preventing the "black box" effect of autonomous agents.

How it works

The platform uses a Manager-Workers architecture where a central Manager agent orchestrates various Worker agents. It leverages the Matrix protocol (via Tuwunel server and Element client) to create IM rooms where humans, the Manager, and Workers collaborate in real-time.

Key technical components include:

• Higress AI Gateway: Centralizes traffic and secures API credentials so Workers never see raw keys.
• MinIO Shared File System: Allows agents to exchange files and information without consuming excessive LLM tokens.
• Kubernetes-native Control Plane: Enables declarative resource management of teams and workers via YAML.
• Multi-Runtime Support: Allows different runtimes (e.g., QwenPaw, Hermes, OpenClaw) to coexist in the same room.

Who it’s for

It is designed for enterprise users and developers who need to deploy collaborative AI agent teams with full human-in-the-loop capabilities, strict credential security, and the ability to host the entire stack privately.

Highlights

• Human-in-the-Loop: All interactions occur in a visible IM room, allowing humans to intervene at any time.
• Enterprise Security: API keys are stored in the gateway, not within the agent containers.
• Zero Configuration IM: Includes a built-in Matrix server for immediate onboarding without external bot approvals.
• Skills Ecosystem: Workers can pull from a library of over 80,000 community skills via skills.sh.
• K8s Native: Full support for deployment via Helm charts on Kubernetes.