agentscope-ai/AgentTeams

What it solves

AgentTeams is a collaborative multi-agent runtime platform designed for enterprise environments. It solves the problem of managing multiple AI agents in a way that is secure, auditable, and transparent, allowing humans to monitor and intervene in agent-to-agent and human-to-agent collaborations without relying on "black box" processes.

How it works

The platform uses a Manager-Workers architecture where a central Manager agent orchestrates various Worker agents. It integrates several key components to facilitate this:

• IM-based Collaboration: Uses the Matrix protocol (via Tuwunel server and Element client) to create shared "rooms" where humans, the Manager, and Workers interact.
• Multi-Runtime Support: Allows different agent runtimes (like OpenClaw, QwenPaw, and Hermes) to coexist and collaborate in the same room, leveraging the specific strengths of each.
• Secure Gateway: Employs the Higress AI Gateway to centralize traffic and keep sensitive API credentials hidden from the Worker agents.
• Shared Storage: Uses MinIO as a shared file system for agents to exchange information efficiently, reducing token usage.
• Kubernetes-Native: Offers a Kubernetes-native control plane for declarative resource management of Workers and Teams.

Who it’s for

It is intended for enterprise users and developers who need to deploy private, secure, and human-in-the-loop multi-agent systems for complex tasks, particularly those requiring high visibility and auditing capabilities.

Highlights

• Human-in-the-Loop: Full visibility and intervention capabilities within the IM environment.
• Enterprise Security: API keys and credentials stay in the gateway, not with the Worker agents.
• Privacy: Based on the decentralized Matrix protocol, allowing for self-hosting and no vendor lock-in.
• Rapid Deployment: One-command setup for local environments or Helm charts for Kubernetes deployments.
• Extensible Skills: Workers can pull from a community ecosystem of over 80,000 skills via skills.sh.