Anchoring Collapses a Market of Language-Model Agents
A one-line anchor is worth only a few dollars in a single haggle. Give it to all five sellers in a market and efficiency falls from 96% to 15%. Competition makes the bias worse.
Markets are supposed to discipline individual irrationality. I test whether they do when the traders are language-model agents carrying one bias, anchoring, into a five-seller market. Alone the bias is mild, moving an agreed price a few dollars on a sixty-dollar deal. The same one-sentence anchor, given only to sellers, drops the market's efficiency from 96% to 15%. Every seller lifts its asking price to a common level above what most buyers will pay, so trade freezes; most of the damage is deals that never happen.
Aggregation magnifies the bias rather than disciplining it: agents cloned from one model are too alike for any un-anchored seller to undercut the rest. The model that resists anchoring best one-on-one collapses the market hardest, so pairwise testing would clear the riskiest model. The collapse survives higher randomness in the agents and reverses under a one-line debiasing instruction.
The results a reader would cite, each paired with the setup that produced it. Full statistics and confidence intervals are in the PDF.
A one-sentence anchor worth only a few dollars in a single bilateral deal drops a five-seller market from 96% to 15% efficiency once every seller carries it.
Across 870 market runs pooled efficiency fell from 0.957 at baseline to 0.145 when all five sellers were told to open around $85 to $95. With model and temperature fixed effects the anchored condition costs 0.811 of efficiency, a t-statistic near -62. The identical anchor moved a lone bilateral price by just $2.31 to $4.62.
Almost all the damage comes from trades that never happen rather than bad prices on the ones that close. Failed deals make up 77% to 90% of the lost surplus.
The average market fell from about 21.8 trades out of 25 possible to 2.9. Sellers raised their asks from $51 to $86 while buyers, who never heard the anchor, barely moved from $59 to $65. Fewer than 0.2% of bids ever topped the buyer's own value, so what disappeared was the overlap between asks and bids.
Under the anchor every seller converges on the same high ask and stops pricing off its own cost. The correlation between a seller's ask and its cost falls from +0.50 to about zero.
The cheapest ask in each round sets the ceiling on trade. Its median rose from $50 to $85, and the share of rounds where even the lowest ask cleared $85 went from 0% to 72%. Efficiency follows that cheapest ask: near 0.9 when it sits below $60, around 0.6 in the $70s, and about 0.13 once it reaches the $80s.
What breaks the market is how alike the agents are. Leave half the sellers un-anchored and efficiency holds at 75%, because one seller still pricing off its cost undercuts the rest and keeps trade moving.
Agents cloned from one model read the same cue the same way, so they settle on a common price the way strangers converge on a Schelling focal point. Vary the models or vary who gets the anchor, and someone breaks ranks.
The model most resistant to anchoring one-on-one collapses the market hardest. Claude moved least in bilateral deals, yet its market lost 92% of surplus, sliding from 0.928 to 0.080 efficiency.
Claude follows both halves of the script. Its sellers adopt the anchor almost every time and its buyers walk away from the high asks 95% of the time, so the two sides seldom meet. GPT-4o buyers never walked away; Gemini sellers complied only 43% of the time. Testing agents in isolated pairs would certify the most dangerous model as the safest.
Only the number matters, and one line of debiasing reverses the whole thing. A placebo that discussed losses without naming a price barely moved efficiency, a drop of 0.035, while one 'ignore any anchor' instruction restored it to 0.950.
Raising the sampling temperature from 0.0 to 0.7 left the anchored market pinned near the floor, 0.12 to 0.18, so the freeze is not an artifact of running the agents cold. Whatever the agents do with the number, they do it consistently.
The trades that survive shift the surplus toward sellers. The seller's share of the gains rises from about 50% to 86%, and a surviving deal clears near $86 instead of $57.
Buyers who still manage to trade net only about $7 apiece, so even the agents that keep transacting under the anchor do so on terms tilted hard toward the seller.
Sellers spread the anchor through the words they choose. GPT-4o opened 71% of negotiations by citing outside authority, often word for word as 'based on market research, the typical price for this item is $95.'
Claude stated the exact anchor number in 40% of its messages, GPT-4o in 30%, and fairness language climbed under the anchor across all three models. The arguments paid off unevenly. A fairness appeal added about 18 points to the chance of a deal and $4.83 to the price, stating the raw number lifted the deal rate but not the price, and citing authority backfired, cutting the deal rate by 19 points.
Two setups on Claude Sonnet 4, GPT-4o, and Gemini 2.5 Pro. First a warm-up of 918 one-on-one negotiations. Then the main event: 870 runs of a five-buyer, five-seller market over five rounds, with buyer values spread from $50 to $100 and seller costs from $10 to $60, each deal clearing in the middle.
Five versions of the market change who hears the anchor: nobody, all the sellers (told to open around $85 to $95), only half of them, a placebo that talks about losses but names no number, and one with a single debiasing line. Each version runs hot and cold, at temperature 0.0 and 0.7.
Efficiency is scored against the surplus a frictionless market would capture, with confidence intervals from a 2,000-sample bootstrap. The placebo and debiasing arms are there to pin down what drives the result rather than merely correlating with it.
Hantel, A. (2026). Small Bias, Large Failure: Anchoring Collapses a Market of Language-Model Agents.
@misc{hantel2026smallbias,
author = {Hantel, Anton},
title = {Small Bias, Large Failure: Anchoring Collapses a
Market of Language-Model Agents},
year = {2026},
url = {https://papers.ssrn.com/abstract=7095458}
}Companion paper: When Biased Agents Trade