The University of Hong Kong's Data Intelligence Lab (HKUDS) has open-sourced AgentSpace, a multi-agent collaboration platform. To break free from the limitations of single-user terminals and scattered conversations, AgentSpace defines AI agents as "digital employees," offering team workspaces with multi-role permissions, workflow coordination, and audit monitoring.
At its core, the standardized routing layer AgentRouter can dispatch the same agent to different runtimes — like Claude Code, Codex CLI, OpenClaw, or Hermes Agent — without altering the business queue. It unifies runtime status, execution results, and error diagnostics. A companion digital-employee dashboard supports agent transfer requests and provides skill binding and tool approval channels for owners and administrators.
Security and access control span workspace members, channel access, private chat privacy, remote daemons, and more. AgentSpace also integrates a Google Workspace OAuth authorization proxy for agents, enabling them to autonomously create or read and write cloud documents — and automatically trigger approval workflows when permissions are missing.
The roadmap ahead includes more robust AgentRouter session management, multi-agent sandbox isolation strategies, and a runtime tool marketplace.