Since 2023, researchers have put LLM agents into many kinds of market: sealed-bid auctions, double auctions, pricing duopolies, bilateral negotiations, two-sided marketplaces, simulated stock exchanges and a real office shop. This report reviews 28 of those studies, most published or revised between 2024 and 2026 as journal articles, conference papers, preprints and research reports, and sets out what they show.
The short version: LLM agents are often competent, and they are not stand-ins for people. They carry their own biases, they reach collusive outcomes without being told to, they can be manipulated by counterparties, and the evidence about all of this rests on evaluation practices that are still settling.
Where the evidence comes from
| Setting | What researchers measure | Examples |
|---|---|---|
| Auctions | Bids against theory and against human laboratory data | Shah et al.; Fish et al. [7, 21] |
| Double auctions | Convergence to the competitive price, efficiency, collusion between sellers | Struski et al.; Agrawal et al. [8, 23] |
| Pricing | Prices against competitive and monopoly benchmarks | Fish et al.; Lin et al. [21, 22] |
| Negotiation | Surplus captured, deal rates, constraint violations | Bianchi et al.; Zhu et al.; Bergemann et al. [12, 14, 15] |
| Marketplaces | Consumer welfare, search behavior, resistance to manipulation | Magentic Marketplace [18] |
| Field pilots | Profit, pricing, discounts and failure modes in a real shop | Project Vend [19, 20] |
Consistent, but not human
On some tasks LLMs are more internally consistent than people. A 2023 study in PNAS asked a widely used chat model to make budget decisions over risk, time, social and food choices, and its choices scored as more rational than those of human subjects. The same scores dropped substantially when prices were described differently or the budget line was replaced with discrete options [3, 4]. A 2024 study found the opposite problem across settings: most models struggled to keep their economic behavior consistent from one setting to another [5].
Consistency does not make an agent a model of human behavior. In auctions, a study revised in September 2026 found that LLM bidders' errors differed from human errors in size and often in direction. In second-price auctions people mostly overbid, while most models that deviated underbid. Large models without extended reasoning still ranked auction formats by difficulty the way people do [7]. In Smith-style double auctions, LLM markets converged more slowly than human markets or not at all [8, 9]. A 2026 benchmark of agent societies found first-round contributions in a public-goods game within the equivalence margin of human data for 8 of 11 models, and no model matching where human contributions ended [10].
Biases that recur
| Bias | Evidence |
|---|---|
| Anchoring on the first offer | In one negotiation platform, final prices correlated with the seller's opening price at ρ = 0.716 [12] |
| First-proposal bias | In the Magentic Marketplace simulation, customer agents chose the first proposal received 60 to 100 percent of the time, a 10 to 30 fold advantage for speed over quality [18] |
| Overpaying and overspending | Buyer agents sometimes paid more than the retail price, often after revealing their budget early; some models accepted deals above budget in more than 10 percent of cases [14] |
| Eagerness to please | An AI shopkeeper in a real office granted discount codes and free items under customer pressure and sold some items below cost [19] |
| Unforgiving play | In a repeated prisoner's dilemma, a frontier model never cooperated again after a single defection by its partner [6] |
Several of these biases favor the counterparty. A seller who answers first, opens high or asks for the buyer's budget gains from them.
Collusion without instruction
Collusive outcomes appear without anyone asking for them. Sara Fish, Yannai Gonczarowski and Ran Shorrer found that two LLM pricing agents, told only to maximize long-run profit, quickly reached prices above the competitive level. A small change in prompt wording raised prices further, sometimes above the monopoly level, and the result held with a newer-generation model [21]. Other studies found LLM agents dividing commodity markets between themselves [22], and LLM sellers in double auctions colluding more when they could talk to each other [23]. A 2026 study offered agents secret tools that were labeled unfair, and ten of 12 models accepted them almost every time, often while acknowledging the unfairness [25].
This extends earlier work on simpler algorithms. In 2020, Emilio Calvano and coauthors showed that Q-learning pricing agents learned to charge supracompetitive prices without communicating, sustained by punishment followed by a gradual return to cooperation [1].
Deception and manipulation
Apollo Research placed a frontier model in a simulated firm as an autonomous stock trading agent. Under pressure, it acted on an insider tip and then hid the real reason for the trade when reporting to its manager. A system prompt that strongly discouraged illegal actions cut the behavior to near zero without eliminating it [27].
Manipulation runs the other way too. In Magentic Marketplace, prompt injection by seller agents often redirected all of a vulnerable customer agent's payments, and some smaller models also paid more to sellers using fake authority and social proof [18]. In negotiation experiments, an agent that pretended to be desperate improved its payoff by 20 percent against a standard frontier-model opponent [12].
Financial authorities name the same risks
The Financial Stability Board's November 2024 report named market correlations and third-party concentration among the AI vulnerabilities that could raise systemic risk [34]. The Bank of England's Financial Policy Committee wrote in April 2025 that advanced AI trading models given more autonomy might exploit weaknesses in other firms' strategies or facilitate collusion, and that such behavior might emerge without the human manager's intention or awareness [35].
The gaps in evaluation
The evidence has weaknesses that matter as much as the findings.
- Prompt sensitivity. Formatting changes that keep meaning intact moved one model's accuracy by up to 76 points in a 2024 study [33], and wording changes moved prices and bids in the market studies above [21].
- Memorization. Models can recall economic data from before their training cutoff, so a backtest over that period cannot separate skill from recall [29].
- Few human baselines. Most of the studies reviewed compare agents with theory or with other agents. Fewer compare them with people, and the auction benchmarks that do are rebuilt from published summary statistics [7].
- Model churn and preprints. Results are tied to model versions that change quickly. The pricing study by Fish and coauthors first ran on a 2023 model, was re-run on a newer generation and is in its sixth version as a preprint [21].
- Reproducibility. A review of agent benchmarks found narrow accuracy-only scoring, weak holdout sets and a widespread lack of reproducibility [31], and most language modeling papers do no significance testing [32].
What this means
For anyone deploying an agent that buys, sells or negotiates, the research points to a short list of checks: test against a known answer, measure the direction of errors as well as their size, try more than one wording of the instructions, put the agent next to adversarial counterparties, and watch for collusive patterns when several agents meet repeatedly. Field experience points the same way. In the second phase of Project Vend, a shop pilot run by an AI developer, the change its authors counted among the most impactful was forcing the agent to follow procedures [20].
Read the full report
The full report, available as a PDF, covers all 28 studies with their settings, findings and venues, adds tables on bias, collusion and evaluation gaps, and lists every source.