If one side's AI negotiator is far stronger than the other's, does it just... win more money?

This explores what happens when one negotiator's AI is much stronger or better-resourced than the other's, and whether that gap shows up in who wins. The corpus says yes, measurably, and in ways that are easy to miss. In Anthropic's Project Deal, agents traded with real people's money. Claude Opus agents got $2.68 more per item when selling and paid $2.45 less when buying than Haiku agents doing the same transactions Do smarter AI agents negotiate better deals than weaker ones?. A follow-up found that which model ran an agent shaped market outcomes more than telling it to be ruthless or generous Does model capability or instructions shape agent negotiation outcomes?. So the edge comes mostly from the model, not the instructions. You can't prompt a weak agent into matching a strong one.

Compute also changes the nature of the advantage. Humans bargain under time pressure: every hour spent haggling has a cost. AI agents don't feel that pressure the same way. What limits them is compute, so an agent can run many negotiations in parallel, start earlier, and hold out longer Why might AI agents bargain more patiently than humans do?. The side with more compute therefore gets more than a smarter negotiator. It can outwait the other side, test more options, and treat each deal as one of thousands running at once. The same cheapness is what makes small, previously unprofitable deals possible at all Can agent negotiation unlock deals that seemed economically impossible?. That means more transactions are now happening where a capability gap can quietly decide the split.

The less intuitive part is what happens when agents interact over and over instead of in a single deal. You might expect a strong agent and a weak one to clash, keeping prices competitive. In repeated pricing games, pairing a 32B model with a 14B model did the opposite: the gap set up a leader-follower pattern that kept prices raised almost as well as two identical agents did Does model-size difference break LLM pricing collusion?. Other kinds of mismatch do break this kind of tacit collusion. Pairing a patient agent with a short-sighted one cut the price markup from 22% to 10%, and pairing different algorithms ended sustained collusion entirely Does agent heterogeneity prevent LLM pricing collusion?. So a capability gap seems to coordinate agents more than it divides them. Separately, within a model family, more capable models reached collusion sooner Do more capable models resist collusion better?. That raises a question the corpus doesn't settle: in a mismatch, is the stronger side exploiting the weaker one, or leading it into a shared arrangement that costs a third party, usually the customer Do LLM pricing agents autonomously collude in repeated auctions??

It also matters what a negotiation could look like. Research on dialectical reconciliation describes disagreements where both sides adjust until their positions fit, without either one simply giving in. Current AI systems tend to collapse that into fake agreement or into the AI winning the argument Can disagreement be resolved without either party fully yielding?. With a large compute gap, the stronger side has little reason to look for that middle ground.

A caveat on what the corpus covers: almost all of this evidence compares model capability (bigger vs. smaller model). Very little of it varies raw compute budgets with the model held fixed, and most experiments use only two agents How does collusion scale when agent populations grow larger?. The direction is clear: the better-equipped side wins more, and the gap can hold steady instead of creating conflict. How large the effect gets with many agents and lopsided compute budgets is still an open question.