A new multi-agent AI architecture called PANDA lets large groups of independent AI agents assemble themselves into task-specific teams, without a central coordinator.
Researchers built PANDA as a decentralized architecture for multi-agent large language model systems. Instead of relying on a central orchestrator, PANDA lets agents advertise their capabilities, find each other, and form small teams on demand for each task, choosing from three team structures - star, chain, or mesh - depending on what the task needs. The team tested PANDA on HotPotQA, a benchmark that requires pulling together clues from multiple documents to answer multi-step trivia questions. PANDA scaled to thousands of agents, assembled teams in milliseconds, matched the accuracy of existing state-of-the-art systems while using up to 8 times fewer resources, and kept completing 100% of tasks even when parts of the system failed.
Most multi-agent AI frameworks today assume you already know which agents you need and wire them together by hand, which breaks down once you are coordinating thousands of agents across many teams. PANDA's approach looks more like how large distributed systems already work - loosely coupled, self-healing, with trust relationships standing in for central oversight. That is a more realistic model for a future where agents increasingly come from different vendors and have to find each other on the fly, rather than one engineering team hand-assembling a pipeline.
Whether a web-of-trust model holds up as well as the scaling numbers do remains to be seen outside a single benchmark - trusting an unfamiliar agent is a different problem than trusting your own.