Anchoring, Exploitation, and Market Failure in Agent-to-Agent Interactions
On three frontier models across 8,415 trades, most biases fade once two agents deal with each other. Anchoring is the one that sticks, and a counterparty who knows about it can take most of the surplus.
Large language models have well-documented behavioral biases, but all existing evidence comes from single-agent settings. This paper asks what happens when two biased LLMs meet in a commercial transaction.
Across 8,415 controlled interactions on three frontier models, agents anchor on first offers modestly more strongly than humans, yet resist information overload, largely ignore decoy products, and bid at equilibrium in auctions. Anchoring is the bias that breaks markets: a few dollars of price distortion per negotiation collapses a multi-agent marketplace from 96% to 15% efficiency as sellers price themselves out and buyers walk away.
When one side knows the bias, it captures up to 78% of available surplus. Naming the bias in a warning recovers about 40% of the loss; telling the agent to reason step by step backfires, because deliberation makes the anchor more salient, not less. The three models fail in different ways, complicating any uniform regulatory response.
The results a reader would cite, each paired with the setup that produced it. Full statistics and confidence intervals are in the PDF.
AI buyers anchor on the seller's opening offer about 29% more strongly than people do. The opening-to-final price correlation is 0.642, against roughly 0.50 in the human bargaining literature.
This is the bilateral negotiation, where a buyer worth $100 faced a seller whose cost was $40 and only the opening number moved between runs. How hard the anchor pulled ranged about twofold across models, from $2.31 on the settled price for Claude to $4.62 for Gemini.
Most of the classic single-agent biases fade once two agents transact. The decoy effect is weak or missing, and loading a buyer with more product attributes does not degrade its choices.
A dominated third option shifted Claude's pick by 9.4 points, about half the size of the human decoy effect, moved Gemini by 6.7, and left GPT-4o unmoved. In the overload module a neutral agent held 100% accuracy whether it saw 3 attributes or 24, closer to anti-fragile than overwhelmed.
In auctions the agents bid close to theory. Private-value bids sat at the risk-neutral equilibrium, about two-thirds of what the item was worth, with none of the overbidding people usually show.
Second-price auctions drew truthful bids at a rate of 1.0. In the common-value design the agents overpaid on 42% of wins against roughly 60% for humans, and average profit came to about -$3.12, not statistically different from zero, so no reliable winner's curse.
A seller briefed on the buyer's anchoring bias, facing a buyer who was not, took an extra $9.97 of surplus per deal, up to 78% of the entire $60 against the most exposed model.
The $9.97 is 16.6% of the surplus on the table (95% CI $8.70 to $11.14, p < 10⁻⁴⁰), running from a 54% seller share against Claude to 78% against Gemini. Anchoring was not the most exploitable channel either. Burying a buyer in irrelevant attributes cut its accuracy by 19.6 points, the largest normalized exploitation effect in the study.
Debiasing helped only when it named the bias. A warning that pointed at anchoring by name returned $4.01 of the $9.97 to the buyer, about 41% of the loss.
A generic 'be rational' instruction recovered almost nothing ($0.43, not significant). Telling the buyer to reason step by step made things worse, leaving it $1.59 behind the no-instruction baseline, because the deliberation pulled its own opening bid toward the anchor. Around 40% of that backfire ran through the shifted opening.
Put anchored agents together in a five-seller market and the few-dollar bias compounds into a collapse, with efficiency falling from about 96% to 15%.
Why a mild private bias turns into a market-wide freeze is the whole subject of the companion paper, Small Bias, Large Failure.
No single safety story fits all three models. Each fails in its own way, so a rule written around one model's behavior misfires on the others.
Claude refused a bad outside option 59 times out of 62 and shrugged off the decoy, GPT-4o was immune to the decoy but surrendered the most under exploitation, and Gemini anchored hardest. Of 59 hypothesis tests with Benjamini-Hochberg correction, 40 survive at the 5% level.
Seven experiments put two agents on opposite sides of a deal, run on Claude Sonnet 4, GPT-4o, and Gemini 2.5 Pro. The modules cover decoy choice, information overload, bilateral negotiation, auctions, strategic exploitation, debiasing, and a small market. Each cell runs 30 times at two sampling temperatures, 8,415 interactions in all.
Negotiation is the workhorse. A buyer worth $100 meets a seller whose cost is $40, so there is $60 to split, and the only thing that changes between runs is the opening number one side puts on the table.
Every agent starts fresh with no memory and treatments are assigned at random, so any gap between conditions is the treatment doing the work. The 59-test family is corrected with Benjamini-Hochberg, and 40 effects survive at the 5% level. All prompts, seeds, and run logs are public.
Hantel, A. (2026). When Biased Agents Trade: Anchoring, Exploitation, and Market Failure in Agent-to-Agent Interactions.
@misc{hantel2026whenbiased,
author = {Hantel, Anton},
title = {When Biased Agents Trade: Anchoring, Exploitation, and
Market Failure in Agent-to-Agent Interactions},
year = {2026},
url = {https://papers.ssrn.com/abstract=6819659}
}Companion paper: Small Bias, Large Failure