Module 01
Decoy choice
A dominated third option barely swayed the agents. It moved Claude's pick by 9.4 points, about half the human decoy effect, nudged Gemini by 6.7, and left GPT-4o unmoved.
Under review · Journal of Risk and Uncertainty · 2026
When Biased Agents Trade
Anchoring, Exploitation, and Market Failure in Agent-to-Agent Interactions
Anton Hantel · MIT
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.
Abstract
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.
Design and treatments
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.
The seven modules
Seven experiments, two agents each.
Every module puts two agents on opposite sides of a deal and isolates one behavior, from decoy choice through a small five-seller market.
Module 02
Information overload
More product attributes did not degrade choices. A neutral agent held 100% accuracy whether it saw 3 attributes or 24, so the agents were closer to anti-fragile than overwhelmed.
Module 03
Bilateral negotiation
Anchoring is the bias that stuck. The seller's opening offer pulled the final price 1.29 times harder for an AI buyer than for a person, an opening-to-final correlation of 0.642 against 0.497 in human studies, on a $60 surplus split.
Module 04
Auctions
Bidding tracked theory. Private-value bids sat at the risk-neutral equilibrium near two-thirds of value, and second-price bids came in truthful. There was no reliable winner's curse either, with mean profit at -$3.12, not different from zero.
Module 05
Strategic exploitation
A seller briefed on the buyer's bias, against a buyer who was not, took an extra $9.97 per deal, up to 78% of the $60 surplus against the most exposed model. The largest normalized effect came through information, not anchoring: burying a buyer in irrelevant attributes cut its accuracy by 19.6 points.
Module 06
Debiasing
Only a specific warning helped, returning $4.01 to the buyer, about 41% of the loss. A generic 'be rational' did nothing, and step-by-step reasoning trended the other way, leaving the buyer $1.59 worse off (p = 0.067).
Module 07
Multi-agent market
Put anchored agents in a five-seller market and efficiency fell from about 96% to 15%. The companion paper, Small Bias, Large Failure, traces the mechanism.
Synthesis
Across the seven, the three models fail in different ways. Claude refused a bad outside option 59 times out of 62, GPT-4o was immune to the decoy yet surrendered the most under exploitation, and Gemini anchored hardest. Of 59 hypothesis tests with Benjamini-Hochberg correction, 40 survive at the 5% level, so a rule written around one model misfires on the others.
Cite
Hantel, A. (2026). When Biased Agents Trade: Anchoring, Exploitation, and Market Failure in Agent-to-Agent Interactions. SSRN Working Paper 6819659.
@unpublished{hantel2026whenbiased,
author = {Hantel, Anton},
title = {When Biased Agents Trade: Anchoring, Exploitation, and
Market Failure in Agent-to-Agent Interactions},
year = {2026},
note = {SSRN Working Paper 6819659},
url = {https://papers.ssrn.com/abstract=6819659}
}JEL classification D91, D18, D44, K24, L86
Keywords agentic commerce, agent-to-agent interactions, large language models, behavioral biases, anchoring, winner's curse, decoy effect
Companion paper: Small Bias, Large Failure