AGTBRIEF
Northeastern researchers propose PANDA, a multi agent design that drops the central registry and lets agents self form teams. The 8x benchmark claim is single dataset.
When a multi-agent AI system goes down, the agents themselves are usually the last thing to fail. The first thing to fall over is the bus: the orchestrator or central registry that decides which agent is allowed to talk to which. A new research preprint, PANDA, treats that bus as the actual problem and proposes moving it off the critical path.
Submitted September 29, 2026 by Matthew D. Laws and Cristina Nita-Rotaru at Northeastern University, the paper describes a decentralized design where each agent advertises its own capabilities into a local registry rather than a central repository, and self-forms small specialized teams per task. Agents can belong to multiple teams and schedule work via FIFO or lottery queues, per the paper's HTML. Communication splits between GossipSub for collective broadcasts and TCP for task-team messages.
The authors report up to 8x efficiency over state-of-the-art baselines on HotPotQA, a standard multi-hop question-answering benchmark, and 100% task completion under injected infrastructure and orchestration faults where existing systems reportedly failed. Both numbers are author-reported, single-benchmark, and untested outside the paper's fault model. PANDA is a preprint, not a deployed product. Decentralized routing's untested case is adversarial agents, where the web-of-trust model the authors invoke would have to do the work the central bus no longer does.