Cooperative Ai Agents Using Supervised Global Workspaces

Parham Aarabi, Yiqian Qin

Abstract: Multi-agent systems can tackle tasks surpassing individual agent capabilities by harnessing autonomous, reactive, proactive, and social agents that collaborate, coordinate, and negotiate solutions. We propose a system merging these principles with a global workspace, overseen by a single supervisor agent. Each specialized agent has unique skills and writes to a shared workspace, ensuring synchronized progress. We demonstrate this approach in a e-commerce design assistance recommendation scenario involving multiple agent personas. Observations reveal that strong supervision enforces orderly behaviour, while diverse agent personas foster richer dialogues and improved outcomes. (Abstract)

External IDs:dblp:conf/ieeecai/AarabiQ25