Teams of AIs solve tasks faster — but stumble when everyone's choices collide

Split a big job among several AI assistants and they fly: one drafts, one checks, one edits. That divide-and-conquer trick works because each piece can be solved on its own. But some problems refuse to be sliced up.

Think of two chess players deciding their opening strategy. Your best move depends on what your rival will do — and their best move depends on you. Neither side can plan in isolation. This paper calls that knot "stance entanglement": when every stakeholder's decision hinges on everyone else's, you can't farm out the pieces.

The fix, called Multi-Agent Fictitious Play, gives each stakeholder its own AI agent and lets them circle toward a stable answer. The engine is an old game-theory move: each agent looks at the average of what its rivals have done across all past rounds, then picks the best reply to that mix. Round after round, agents poke holes in each other's plans and patch their own. The decisions stop swinging wildly and settle.

Tested on competitive scenarios where you must lock in a strategy before acting, this beat both one-shot AIs and ones that simply re-ran the same prompt several times — measured by how often it won head-to-head and how well it held up against varied opponents.

The paper only tests strategic, competitive setups, so it's unclear how well this travels to messier real-world negotiations.

Why you should care: Most real decisions — pricing, negotiating, bidding — depend on what the other side does. This shows AI can reason through that back-and-forth instead of pretending each side decides alone.