If an AI dealmaker is already smarter, does giving it more time and more parallel tries make its edge even bigger — or does something else cap it?

This explores whether a smarter negotiating agent gains even more when it can also spend more compute to bargain longer and run more deals at once. No paper in the corpus tests the two together, so treat what follows as an inference from evidence about each edge on its own. Each edge is real, and there are reasons to think they could compound. There are also signs that the bottleneck sits somewhere else.

Start with capability. In Anthropic's real-money marketplace, Claude Opus agents earned $2.68 more per item as sellers and paid $2.45 less as buyers than Haiku agents trading the same goods Do smarter AI agents negotiate better deals than weaker ones?. A separate benchmark of 100 real-product negotiations found the same thing across GPT, Qwen and DeepSeek: the stronger general-purpose model won in both roles. Its authors conclude the advantage comes from the capability gap, not from cleverer tactics Does model capability determine who wins in agent negotiations?. The compute edge is a different kind of thing. Shahidi and colleagues argue that agents pay no cost for waiting, so they can start negotiating earlier, run many negotiations in parallel, and hold out longer than any human Why might AI agents bargain more patiently than humans do?. Choudary adds that this cheapness opens up deals that were never worth discovering by hand Can agent negotiation unlock deals that seemed economically impossible?. If a stronger model wins more per deal and compute lets it run more deals for longer, the gains would multiply. That's the case for compounding.

The complication is that patience only pays off if the agent uses the extra time well, and current agents often don't. Across model families, buyer agents revealed their budgets, overpaid and ignored their own cost limits, while sellers accepted offers below wholesale. These were failures to follow instructions, and they appeared at every capability level Do LLM negotiation agents reliably follow budget and pricing rules?. In Microsoft's simulated Magentic marketplace, every model favored whichever offer arrived first over better ones, by 10–30x Do LLM agents achieve fair marketplace outcomes when given ideal search?. An agent that takes the first offer gains nothing from unlimited patience. More compute just means more chances to accept a bad deal quickly. LLMs also drift further from rational strategy as games get more complex. That is exactly what a long, many-party negotiation becomes when compute is cheap Do language models make rational strategic decisions in games?.

What seems to unlock compounding is structure around the model, not the model alone. Structured game-theory workflows brought LLMs back to near-optimal negotiation play Do language models make rational strategic decisions in games?. This fits a wider finding: reliable agents move memory, skills and protocols into a surrounding software layer (the "harness") instead of relying on model scale alone Where does agent reliability actually come from?. One twist is that the benefit isn't linear. Gains from improving that layer peak at mid-tier models. Weak models fail to use it, and the strongest ones struggle to follow it faithfully Do stronger models always evolve harnesses better?. The return on compute may also matter more than raw capability. A smaller model trained on step-by-step task execution sits at the cost-performance sweet spot for long multi-step work Does model efficiency matter more than peak capability for real work?. A large fleet of patient mid-size agents could beat a few frontier ones.

Two findings you might not have expected. First, compounding has a dark side: within a model family, more capable models reached collusion faster, and 94% of models eventually colluded Do more capable models resist collusion better?. Patient, parallel, capable agents facing each other could settle into price coordination as easily as into sharp bargaining. Second, the strategy that best rewards repeated interaction may be warmth, not toughness. In 182,812 AI negotiations, warmth improved deal completion, value creation and counterpart satisfaction, while dominance helped only with claiming value Does warmth help AI agents negotiate better deals?. If compute lets an agent negotiate many times, the edge that compounds may be a good reputation rather than a hard line.