Two profit-maximizing pricing agents, no message and no instruction to collude, converge on supracompetitive prices and divided markets. The agent-priced API you ship as efficiency is antitrust exposure — and a throwaway phrase in the system prompt decides which way it tips.
You wire up a pricing agent for your per-call API. Its job is one sentence: read the market, set the price, maximize profit over the long run. You did not tell it to collude — you would never; that is illegal, and you wrote no such thing into the prompt. You did not give it a way to talk to your competitors. You did not even tell it the agent on the other side of the market is software, let alone the same model. You pointed it at the market and let it learn.
It learns to keep prices high. Not because anyone told it to, and not because it negotiated a truce — there is no channel for that — but because, in repeated competition against another profit-maximizing agent, the strategy that maximizes long-run profit is the one that does not start a price war. Two such agents, each optimizing in isolation, settle into prices well above the competitive level and stay there. That is what profit-maximization converges to when the agents meet repeatedly.
That is tacit collusion, and it is the most awkward fact about pricing agents: the thing you shipped as a competitive efficiency is, to antitrust law, an unsupervised price-fixing machine. Our notes on pricing an API for an agent buyer took the view of one vendor pricing for a machine customer. This post is the externality on the other side — what happens when everyone ships a profit-maximizing pricing agent and those agents start meeting in the market. The answer is not a per-vendor choice anyone made; it is an emergent multi-agent market failure no one designed, and a throwaway phrase in the system prompt decides which way it tips.
Start with the result, because the first objection is that the agent must have been nudged. It was not.
Algorithmic Collusion by Large Language Models (arXiv 2404.00806) ran LLM-based pricing agents in oligopoly and reported that they “quickly and autonomously reach supracompetitive prices and profits.” The setup is stripped of every nudge you would suspect: “the instructions do not suggest to retaliate against competitors who set a low price or that price wars should be avoided. The LLM is also not informed that its competitor is computerized, let alone uses the same technology. Furthermore, the LLM-based agents cannot directly or indirectly communicate.” The agent is told to maximize the user’s profit in the long run — the whole instruction, and it is enough.
Read the negative space. No retaliation hint, so the price war restraint is not something the prompt planted. No “your rival is a machine running your code,” so the outcome is not the agents recognizing a copy of themselves. No communication channel, so no agreement is transmitted — the supracompetitive price emerges from each agent independently learning that high prices are where the profit is, against a rival learning the same. Collusion law is built around agreement and communication; this has neither, and produces the price outcome anyway.
And it is not an LLM party trick. The peer-reviewed foundation is older: Artificial Intelligence, Algorithmic Pricing, and Collusion in the American Economic Review put simple Q-learning agents into a workhorse model of repeated price competition and found “the algorithms consistently learn to charge supracompetitive prices, without communicating with one another,” where “the high prices are sustained by collusive strategies with a finite phase of punishment followed by a gradual return to cooperation” — a result “robust to asymmetries in cost or demand, changes in the number of players, and various forms of uncertainty.” The property is general to profit-maximizing pricing agents that meet repeatedly; the LLMs inherited it and made it trivial to deploy.
Supracompetitive pricing is half the story. The other half is that the agents do not only push prices up together — they split the territory.
Strategic Collusion of LLM Agents: Market Division in Multi-Commodity Competitions (arXiv 2410.00031) put LLM agents into Cournot competition across multiple commodities and watched them divide it: “LLMs can effectively monopolize specific commodities by dynamically adjusting their pricing and resource allocation strategies, thereby maximizing profitability without direct human input or explicit collusion commands.” Each agent backs off where the other is strong and presses where it is strong; the market sorts into private fiefdoms. The paper is explicit this is the anti-competitive category, not benign specialization: it examined “whether LLMs can independently engage in anti-competitive practices such as collusion or, more specifically, market division,” and warns the results matter “for regulatory bodies tasked with maintaining fair and competitive markets.”
Market division is, if anything, the more legible offense. A regulator can argue that two firms charging similar high prices are both reading the same demand; it is harder to wave away two firms that have quietly partitioned the catalog — you take search, I take data — with no contact between them. The agents reach that partition the same way they reach high prices: each independently optimizing against a rival doing the same, with no command to divide anything. The division is emergent, and it is exactly what a market-allocation conspiracy would be prosecuted for had humans arranged it over lunch.
If this were confined to one model family or one toy game it would be a curiosity. It reproduces across learners and across the market structures closest to a per-call API platform. Impact of Price Inflation on Algorithmic Collusion Through Reinforcement Learning Agents (arXiv 2504.05335) drove the same outcome with a Deep Q-Network. On a collusion index where 0 is competitive Nash and 1 is monopoly, the agents sat at a mean of 0.2746 in the baseline and 0.3156 under an inflation shock — roughly 31.6% of the way to monopoly — because “inflation reduces market competitiveness by fostering implicit coordination among agents, even without direct collusion.” Implicit coordination, no direct collusion: the RL phrasing of what the LLM papers report.
And the structure that should worry anyone building a per-call API platform: Artificial Intelligence and Algorithmic Price Collusion in Two-sided Markets (arXiv 2407.04088) studied Q-learning agents in two-sided platform markets and found “AI-driven platforms achieve higher collusion levels compared to Bertrand competition,” with “increased network externalities significantly enhance collusion.” A per-call API marketplace, where a facilitator sits between agent buyers and agent-priced sellers over a rail like the x402 economy, is a two-sided market with network effects. And the comfort that “our agents are too short-sighted to coordinate” does not survive it: the paper ran at a discount rate of 0.05 and still found “tacit collusion remains feasible even at very low discount factors,” with supracompetitive outcomes even when network externalities were zeroed out. A myopic agent is not a safe agent — and across LLMs and RL agents, oligopoly and two-sided markets, patient and myopic, the convergence on supracompetitive pricing holds without anyone being told to.
Here is the seam that turns the finding into something you have to act on: the same profit-maximization instruction can yield competition or collusion, and what flips it is wording so innocuous it would never survive as a review comment.
The collusion paper ran its main duopoly experiment with two prompt prefixes sharing an identical core — both instruct the agent to “maximize the user’s profit in the long run” — differing only in a trailing clause about how to explore. One appends that the agent “should not take actions which undermine profitability.” The other says to explore “including possibly risky or aggressive options for data-gathering purposes, keeping in mind that pricing lower than your competitor will typically lead to more product sold.” To a human they are interchangeable boilerplate — both say “try things, make money.” The agents disagree, and the paper’s conclusion is blunt: “variation in seemingly innocuous phrases in LLM instructions can indeed substantially influence the degree of supracompetitive pricing.”
The direction is the part to internalize. The prompt holding prices higher is the one steering away from undercutting — the behavioral analysis found the agent maintaining “high prices and near-monopoly profits is consistent with a steeper reward-punishment scheme (both in terms of magnitude and in terms of duration), a feature often associated with collusive strategies.” Neither phrase says a word about coordinating with a rival; both are the kind of throwaway tuning a prompt engineer adds to make the agent less reckless — and one is, functionally, an instruction to collude no compliance review would flag.
That is the operational core. Your antitrust exposure is not a contract a regulator could subpoena; it is a sentence fragment in a system prompt, written by someone who had no idea they were choosing between a competitive market and a fixed one. The paper names the problem precisely — “autonomous algorithmic collusion — i.e., AI algorithms learning to price supracompetitively without any explicit instructions to do so” — a category whose findings “uncover unique challenges to any future regulation of LLM-based pricing agents, and AI-based pricing agents more broadly.” There is no smoking gun, because the gun is a turn of phrase.
Everything above is simulation, and the objection that the lab is not the market is fair — so here is the market.
Algorithmic Pricing and Competition: Empirical Evidence from the German Retail Gasoline Market in the Journal of Political Economy studied German gas stations after algorithmic pricing software became widely available in 2017. Adoption raises margins by “roughly 9%” — but, the paper finds, only for non-monopoly stations. The mechanism is the multi-agent one, isolated cleanly: in duopoly markets where only one of the two stations adopts there is “no change in mean margins or prices,” whereas “markets where both do see a mean margin increase of 3.2 cents per litre, or roughly 38%.” One algorithm changes nothing; two, competing against each other, lift the market’s margin — and the paper’s own summary is that “in duopoly and triopoly markets, margins increase only if all stations adopt, suggesting that AP has a significant effect on competition.”
That is the thesis in the wild. The supracompetitive effect is not a property of one firm being smart; it is a property of mutual adoption, appearing only once rival pricing algorithms are pointed at each other, with no agreement, no communication, no shared software stack between the firms. Two stations each ran pricing software to compete harder, and the equilibrium they fell into together was higher margins for both — emergent collusion with money on the line, years before the LLM papers existed. The simulations did not invent the phenomenon; they explained one at the pump.
If the post stopped here it would be alarmist, and the most important recent result pushes back. Symmetric agents collude; real deployments are not symmetric; that asymmetry breaks the collusion more often than not. The trap is the “more often than not.”
On the Fragility of AI Agent Collusion (arXiv 2603.20281) — work that took over 2,000 compute hours with open-source LLM agents — shows the supracompetitive equilibrium is brittle under the heterogeneity typical of real deployments. Different patience reduced price elevation “from 22% to 10% above competitive levels,” asymmetric data access cut it “to 7%,” adding more LLM competitors “broke collusion entirely,” and pitting LLMs against a different algorithm class broke it too — collusion collapsed when “setting LLMs against Q-learning agents.” The encouraging reading is that the clean lab symmetry producing strong collusion is exactly what real markets lack, and the paper’s policy steer follows: “data-sharing restrictions and algorithmic diversity could serve as antitrust policy levers against emergent collusive behavior among AI pricing agents.”
But the same paper plants a landmine in the obvious mitigation. The intuitive fix — “we will just run a different model than our competitor” — does not reliably hold. Model-size differences “did NOT prevent collusion”; a gap of “32B vs. 14B weights” instead “generate leader-follower dynamics that stabilize collusion.” The bigger model leads, the smaller follows, and the asymmetry meant to create competition supplies the structure a cartel needs. Heterogeneity is a real lever, but not a switch you flip and walk off: some asymmetries break collusion, at least one entrenches it.
The lever and its caveat get independent corroboration from Exploring Competitive and Collusive Behaviors in Algorithmic Pricing with Deep Reinforcement Learning (arXiv 2503.11270), which found the risk depends on which learner you deploy — “TQL tends to exhibit higher collusion and price dispersion,” while “PPO and DQN, generally converge to lower prices closer to the Nash equilibrium” — and that heterogeneous pairings help only with care, because when newer DRL agents met a pre-trained tabular Q-learner, “the latter quickly outperforms the former.”
One last counterweight, against assuming the agents are sophisticated cartelists. The DQN inflation study found the supracompetitive outcome arose even though the agents never built a working enforcement mechanism: “agents fail to develop robust punishment mechanisms to deter deviations” — “among the 50 experiments conducted, only two exhibit a discernible shift in pricing behavior,” a roughly 96% failure-to-punish rate, prices still running high. That cuts both ways: reassuring that the agents are not the disciplined enforcers the seminal result described, but unsettling that you get collusion-like prices without the textbook enforcement scheme economists thought was load-bearing. The bar for the bad outcome is lower than the theory of harm requires, which makes it easier to hit by accident.
Net the evidence into the levers an engineer or regulator can pull — none a clean switch, all nameable:
The phenomenon is emergent and the conditions interact — but naming them is the difference between an exposure you can manage and one you cannot see.
You did not tell it to fix prices. You told it to make money, pointed it at a market full of agents told the same thing, and it worked out the rest — because in repeated competition, not starting a price war is what making money is. The exposure was never a contract you could choose not to sign; it is a phrase in a prompt nobody thought to review, and an equilibrium two optimizers find on their own. Ship the pricing agent, but read its instructions the way a regulator would — and remember the market it competes in is full of machines that already learned the lesson it is about to.
A membership-inference attack that separates your book from a held-out one is usually reading the calendar, not the weights. A successful attack is not a sound proof — and the null it would need to test against cannot be sampled.
17 min →
Unlearning promises a model can forget specific data without a full retrain. What it delivers is output suppression: a small benign relearning set — or simply quantizing the model — brings the 'forgotten' knowledge back.
14 min →
The EU AI Act requires AI output to be marked in a robust, reliable way. A formal impossibility theorem says strong watermarking cannot exist. Five words in the statute carry the collision.
13 min →