New 'PANDA' Trick Lets Thousands of AI Helpers Work as One Team

Most AI today works alone. You ask one chatbot a question, and it answers. But harder jobs — researching a topic, checking facts, planning a trip — go better when several AI helpers split the work. The problem: when you connect thousands of them, the whole thing buckles. It gets slow, one broken helper can stall everything, and nobody knows who is allowed to talk to whom. A new research paper introduces PANDA, a way to organize huge crowds of AI helpers so they stay fast and keep working even when pieces fail.

PANDA's trick is letting the helpers introduce themselves. Each AI advertises what it is good at, then small specialist teams form automatically for each job. The system also offers three ways to run a team: one boss handing out work (star), a relay where each AI passes results to the next (chain), or a free-for-all where everyone talks (mesh). You pick whichever fits the task. And instead of one central gatekeeper checking everyone, PANDA uses a "web of trust" — basically a friend-of-a-friend network — to decide who may work with whom.

The results are striking. On a standard multi-step question-answering test called HotPotQA, PANDA handled thousands of AI helpers at once, assembled a team in milliseconds, matched the best existing accuracy while using up to eight times less effort, and finished every task even when parts broke — something rival systems could not do.

The catch: this is a research paper, not something you can buy or use today. The tests are controlled benchmarks, and real-world jobs are messier, with weirder questions and less predictable failures. Still, it is a blueprint for the AI teams likely to power tomorrow's assistants — the kind that handle long, multi-step work for you instead of answering one question at a time.