On the Fragility of AI Agent Collusion

Abstract Recent work shows that pricing with symmetric LLM agents leads to algorithmic collusion. We show that collusion is fragile under the heterogeneity typical of real deployments. In a stylized repeated-pricing model, heterogeneity in patience or data access reduces the set of collusive equilibria. Experiments with open-source LLM agents (totaling over 2,000 compute hours) align with these predictions: patience heterogeneity reduces price lift from 22% to 10% above competitive levels; asymmetric data access, to 7%. Increasing the number of competing LLMs breaks up collusion; so does cross-algorithm heterogeneity, that is, setting LLMs against Q-learning agents. But model-size differences (e.g., 32B vs. 14B weights) do not; they generate leader-follower dynamics that stabilize collusion. We discuss antitrust implications, such as enforcement actions restricting data-sharing and policies promoting algorithmic diversity.

Introduction. Firms are increasingly delegating pricing decisions to algorithms, making algorithmic collusion a central concern for antitrust authorities (Gans 2023). Large language models (LLMs) push this trend further by enabling agents that interpret context and revise strategy autonomously.

For example, Anthropic’s Project Vend gave Claude price-setting authority in a small but realworld retail setting (Anthropic 2025). On a larger scale, Delta Airlines announced that it is using generative AI to determine a portion of its domestic ticket prices (McCarthy 2025). Against this backdrop, a senior U.S. Department of Justice official recently warned this trend could lead to “fully automated cartels operating without any human involvement” (Kavanagh and Magloughlin 2025).

∗Keppo (keppo@nus.edu.sg): National University of Singapore; Li (yuzeli@bu.edu), Tsoukalas (gerryt@bu.edu), Yuan (nyuan01@bu.edu): Boston University. †First version: September 25, 2025 A growing literature shows that autonomous pricing algorithms can sustain tacit collusion without explicit communication (Calvano et al. 2020, Klein 2021, Banchio and Mantegazza 2022).

More recently, Fish et al. (2025) show that LLM-based agents also collude, may converge faster than reinforcement learning (RL) approaches, and do so robustly across prompt variations. Yet existing studies typically assume identical or near-identical agents competing in highly stylized markets. Real deployments are messier: firms adopt models that have different capabilities, different algorithms, proprietary prompts, varying update frequencies, and asymmetric information sets (Huang and Zhu 2023). This raises a central question for competition policy: does LLM-mediated algorithmic collusion survive the heterogeneity typical of real-world deployments, and if variation disrupts collusion, which dimensions matter most? This paper provides systematic evidence.

Related work. The two closest papers are Calvano et al. (2020) and Fish et al. (2025). Calvano et al. (2020) show that homogeneous Q-learning agents can learn tacitly collusive pricing, and they document robustness to standard changes in the environment, including cost and demand asymmetries, stochastic demand, and changes in the number of firms. Fish et al. (2025) show the corresponding result for LLM pricing agents. They compare prompt variants and show that while wording can affect the level of price elevation, supracompetitive outcomes persist; they also report robustness to demand-side asymmetry (unequal product qualities) and stochastic demand.

Our contribution is different. Even if collusion is relatively robust to prompt wording, real deployments differ along other dimensions, and these differences need not be benign. We therefore purposely hold fixed the baseline prompt structure and ask whether tacit collusion survives cross-firm and cross-algorithm differences. We find that several of these heterogeneities suffice to materially weaken or break sustained price elevation even without introducing additional prompt variability.

More broadly, our work connects three literatures: algorithmic collusion, firm heterogeneity, and platform governance.

On algorithmic collusion, a broader literature confirms that reinforcement learning agents generate supra-competitive outcomes (Klein 2021, Hansen et al. 2021, Brown and MacKay 2025, Musolff 2025), with mechanisms including spontaneous coupling (Banchio and Mantegazza 2022).1 However, this literature also reveals sensitivity to implementation details: even minor differences such as exploration rules can shift outcomes (Asker et al. 2022, Abada and Lambin 2023, Wang et al.

2023). This sensitivity motivates our focus on heterogeneity. Meanwhile, classical reinforcement learning faces scalability constraints (Deng 2024, den Boer et al. 2024), driving practical interest toward LLMs, which adapt without retraining (Miller et al. 2024, Tambe 2025, Shetty et al. 2025).

The real-world deployments discussed above (Project Vend, Delta Airlines) exemplify this shift.

On firm heterogeneity, the classical literature offers competing predictions. More competitors increase deviation incentives (Stigler 1964, Green and Porter 1984, Miller and Weinberg 2017, Allon et al. 2019). Cost asymmetries create divergent incentives (Brock and Scheinkman 1985, Athey and Bagwell 2008), size disparities weaken punishment credibility (Genesove and Mullin 2001), and information asymmetries undermine focal equilibria (Laffont and Martimort 1997, Harrington and Skrzypacz 2011). Yet structured heterogeneity can facilitate collusion through focal roles (Athey and Bagwell 2001, Compte et al. 2002, Brown and MacKay 2023).

Method. Research Methodology: Answering this question with field data is currently challenging; LLM pricing deployments are too recent to generate public observations at scale. We therefore adopt a two-pronged approach that leverages economic theory on one hand and laboratory experiments in the spirit of Fish et al. (2025) on the other. The two play complementary roles. Experiments allow us to manipulate dimensions of heterogeneity independently—something observational data, even if available, would rarely permit. But the space of possible variation is large (models differ, prompts differ, information sets differ, etc.). Experimentation without theoretical grounding risks being ad hoc. Theory helps guide our empirical design by identifying which dimensions have equilibrium consequences.

Model: The folk theorem establishes that tacit collusion is sustainable if and only if players are sufficiently patient, making the discount factor a central primitive (Fudenberg and Maskin 1986).

Abreu et al. (1986, 1990) extend this logic to environments with imperfect monitoring, showing that the ability to detect and punish deviations depends critically on the information available to players. These two primitives—patience and information—anchor our initial analysis. We develop a simple infinite-horizon pricing model that allows both to vary across agents. The model yields two comparative statics predictions. First, heterogeneity in patience undermines collusion: when discount factors differ, the impatient agent faces stronger deviation incentives, making coordination on supra-competitive prices harder to sustain. Second, asymmetric access to market data raises the patience threshold required for collusion, because the informationally disadvantaged agent cannot reliably detect deviations and therefore cannot credibly threaten punishment. The model is intentionally stylized; its contribution is not standalone theory but a principled framework linking heterogeneity dimensions to testable predictions. Our approach follows the methodology of behavioral game theory (Camerer 2003)—using equilibrium predictions as benchmarks for evaluating actual behavior without claiming agents compute equilibria explicitly.

We assess collusive outcomes using two criteria: price convergence (whether agents coordinate on stable pricing) and price elevation relative to the static Nash benchmark (the welfare-relevant measure of market power). This dual criterion distinguishes genuine collusion from competitive convergence.

2 Stylized Model This section presents a stylized infinite-horizon pricing model with two sellers. As discussed, the model serves as a qualitative benchmark to generate directional hypotheses, not as a structural description of LLM cognition. The model identifies which dimensions of heterogeneity should matter for collusion under canonical game-theoretic reasoning; our experiments then test whether LLM agents—deployed under realistic conditions without access to the model—produce outcomes directionally consistent with these predictions. Section 2.1 analyzes a baseline case with heterogeneous patience levels, and Section 2.2 extends the model to a heterogeneous data access case.

Table 2: Summary of Key Notation Variable Definition N Number of sellers a Vertical differentiation parameter b Own-price sensitivity parameter d Cross-price sensitivity parameter μ Price elasticity parameter a0 Outside option parameter pt i Price chosen by seller i in period t pt Vector of prices chosen by all N sellers in period t Qi(pt) Demand of seller i in period t δi ∈[0, 1) Discount factor / patience level of seller i ρi ∈[0, 1] Monitoring precision / data access of seller i st ∈{G, B} Monitoring signal in period t c Constant marginal cost πt i Profit of seller i in period t Πi Discounted infinite-horizon profit of seller i Note: Non-bold symbols are scalars; bold symbols are vectors.

2.1 Heterogeneous Patience Level Proposition 1 (Heterogeneous Patience Level Collusion Threshold) For any given price pc ∈ (p∗, pM]:

Discussion. Testing model predictions. Both theoretical predictions find support. Patience heterogeneity reduces collusive outcomes: when two patient agents interact, they converge to prices approximately 22% above the static Nash equilibrium, with pricing rationales that explicitly emphasize avoiding destructive price wars. Setting a patient agent against a myopic one produces meaningfully weaker price elevation, with prices only about 10% above competitive levels. Fully myopic pairs converge essentially to the competitive outcome. Chain-of-thought inspection indicates that the myopic agent’s inability to credibly commit to future punishment weakens the patient agent’s incentive to maintain cooperation. Information asymmetry also undermines cooperation: relative to the homogeneous-information benchmark, heterogeneous data access reduces the deviation from the competitive benchmark from 22% down to 7%, consistent with the model’s prediction that information asymmetry makes tacit coordination more difficult.

Other Sources of Heterogeneity. Cross-algorithm heterogeneity disrupts collusion entirely: setting Q-learning agents against LLM agents prevents sustained coordination. Across runs, they do not reach sustained collusion within 1,000 periods and instead remain near competitive pricing.

The algorithms operate on incompatible signal spaces: Q-learning agents cannot interpret the cooperative pricing patterns that LLM agents attempt to establish, generating instability that drives both agents toward competitive pricing.

Increasing the number of competing agents gradually, from two to five, delays collusion onset and weakens its robustness. With three agents competing, sustained collusion remains feasible and can lead to even higher prices, but coordination/convergence becomes meaningfully slower. With four, prices drop, collusion is further slowed and becomes less stable. With five sellers, sustained collusion fails to emerge within the experimental horizon. These patterns largely align with classical oligopoly theory.

Perhaps most surprisingly, model-scale heterogeneity does not impede collusion. When pairing larger (32B) and smaller (14B) LLM agents (both patient), final price elevation remains close to the homogeneous benchmark, though convergence is substantially slower. Novel dynamics emerge: the larger model assumes a leadership role and anchors the collusive price, while the smaller model adopts a follower position, suggesting that capability asymmetry can facilitate coordination by establishing clear roles.

Finally, additional robustness tests with a closed-source model (GPT-5-mini) serve two purposes: confirming that our main qualitative findings extend to a more capable model, and examining whether LLM agents exhibit sensitivity to explicit collusion-avoidance prompting—a dimension unique to language-model-based agents. Both hold.

Policy Implications: These findings speak to active policy debates that we discuss in detail in Section 7. We provide a brief preview here.

Our patience-heterogeneity results reveal a mechanism distinct from Calvano et al. (2020), who find collusion diminishes when all agents share a lower discount factor: our findings suggest diversity in strategic horizons itself disrupts coordination. Among other implications, this suggests merger reviews should consider whether transactions homogenize pricing objectives (e.g, capital discount rates), not just algorithms (Gal and Rubinfeld 2024).

Our data-access results provide experimental support for the theory underlying the DOJ’s RealPage case (Antitrust Division, Department of Justice 2026), which targeted information-sharing as a mechanism enabling coordination.