Paper explainer
How do we scale AI agents without putting one agent in charge of everything?
Paper: A Responsibility-Centered Coordination Mechanism for Decentralized Multi-Agent Systems
Oct. 2026 · 2 min read · Preprint
Most multi-agent systems rely on a central orchestrator to assign tasks, track progress, and handle failures. But as the team grows, that orchestrator can become a bottleneck.
We've been exploring a different approach in our new paper: A Responsibility-Centered Coordination Mechanism (RCM) for decentralized LLM-based multi-agent systems.
The key idea is simple. When an agent delegates a task, it doesn't give up responsibility for it. It still needs to monitor the work, verify the result, and step in when something goes wrong.
Every agent makes decisions based on its own responsibilities and orchestrates its delegated work locally. Put many of these local orchestrators together, and global coordination emerges without anyone managing the whole system.
The surprising part is that agents figure out how to organize themselves. Some form chains, others become specialist hubs, and some build hybrid networks. None of these structures need to be designed in advance. They emerge as agents work toward a shared goal.
And when things change or fail, agents can adapt and recover locally without waiting for a central controller.
We tested RCM across reasoning, coding, and real executable agent tasks. It outperforms strong centralized and decentralized baselines by up to 18.7 points in task performance and 10.2 points in responsibility, while also improving failure attribution and harmful-action detection, without task-specific training.
Global coordination can emerge from local responsibility.
A Responsibility-Centered Coordination Mechanism for Decentralized Multi-Agent Systems
Thi-Nhung Nguyen, Viet-Thanh Pham, Linhao Luo, Thuy-Trang Vu, and Dinh Phung
Preprint · Paper link coming soon