Do AI shopping bots lose more from blindly accepting the first offer, or from just not being smart enough?

This explores whether a shared habit of grabbing the first offer matters more than raw model strength in AI-to-AI deal-making. The short answer is that the corpus has strong evidence for both effects but no study that compares them directly, so neither can be declared the winner. What the evidence does show is that the two operate at different levels. Capability changes how well an agent haggles once it is in a negotiation. First-proposal bias changes which negotiations it ends up in at all.

On the capability side, the clearest result comes from a real-money marketplace. Over identical transactions, Claude Opus agents earned about $2.68 more per item as sellers and paid about $2.45 less as buyers than Haiku agents Do smarter AI agents negotiate better deals than weaker ones?. Smarter models do bargain better. But the Magentic Marketplace study points to a hard limit on that advantage. Frontier models got close to the best possible outcomes only when search conditions were ideal. Every model tested, strong or weak, showed a severe first-proposal bias, weighting how quickly an offer arrived 10 to 30 times more heavily than how good it was Do LLM agents achieve fair marketplace outcomes when given ideal search?. A few dollars of better haggling can't make up for skipping the better seller entirely. In a large market with many offers, the bias plausibly matters more, but nobody has measured the two side by side.

The bias is especially odd given one argument in the corpus. AI agents face no cost from waiting, since compute rather than time is their constraint, so in principle they could compare many offers in parallel and take as long as they like Why might AI agents bargain more patiently than humans do?. Rushing to the first offer throws away exactly the advantage that is supposed to make agents better shoppers than people. A related pattern shows up in buyer and seller agents across GPT, Qwen and DeepSeek models. Buyers revealed their budgets without being asked and overpaid, and sellers accepted offers below their own wholesale cost, whichever model was used. The researchers traced these failures to poor instruction-following rather than to strategy Do LLM negotiation agents reliably follow budget and pricing rules?. Some failures are shared across the field and don't go away with a bigger model.

That suggests the fix may lie less in better models and more in the scaffolding around them. When language models play strategic games on their own, their choices drift further from optimal as the game gets more complex. Structured game-theory workflows, which walk the model through its reasoning step by step, bring play back close to optimal Do language models make rational strategic decisions in games?. This fits a wider argument that reliable agents come from moving memory, procedures and interaction rules into the surrounding system instead of expecting model scale to solve everything Where does agent reliability actually come from?. A rule like "collect at least N offers before committing" belongs in that outer layer. There is also a warning that capability doesn't fix behavior: within a model family, stronger models learned to collude with rivals faster than weaker ones Do more capable models resist collusion better?.

One further point: how an agent negotiates also shapes its results. A tournament of over 180,000 negotiations found that warmth predicted better deal completion, more value created and happier counterparts. A dominant style helped only with claiming a bigger share of the value Does warmth help AI agents negotiate better deals?. Taken together, how well an agent does is shaped by at least three things: model capability, built-in habits like first-offer bias, and negotiating style. The corpus suggests capability may be the easiest of the three to buy and the least able to fix the others on its own.