If bidding algorithms repeatedly interact, their response to a rival’s price cut can tell us more than a high market price on its own.
An August 27 preprint by Jakub Seredyński and Georgios Tsaousoglou studies simulated electricity markets with reinforcement-learning agents. It screens 18 environments and behaviorally examines three selected cases. Two show punishment-and-recovery responses to deviations; the third lacks comparable punishment. Training uses centralized critics, while bidding actors execute without communication. This is simulation evidence, not a documented real-market cartel. Preprint, methods and results.
Why repeated interaction changes the question
Consider a simplified, invented market in which a few suppliers sell the same service every day. A supplier might earn more today by cutting its price and attracting additional demand. If rivals respond with aggressive price cuts tomorrow, however, that short-lived gain could be outweighed by lower future earnings.
A learning system can discover such a tradeoff through its reward history. No sentence saying “agree on a price” is logically necessary for a strategy that depends on expected rival reactions. Whether the behavior actually emerges depends on the game, the observations, the learning procedure, and how much future reward matters.
This example explains the economic mechanism; it is not a description of human-like intentions inside a model. The cited study concerns reinforcement-learning bidding policies. It does not establish that a conversational assistant would negotiate the same arrangement, or that separately developed commercial algorithms would behave identically.
High prices have several possible causes
Imagine that a simulated transmission bottleneck makes one local generator difficult to replace. Its ability to demand a high price could persist even if every rival ignored its previous actions. A coordinated strategy is a different explanation: maintaining the outcome depends on how participants respond to one another over time.
To distinguish these explanations, an investigator needs a comparison and an intervention. Hold demand, costs, and capacity fixed while changing an agent’s behavior. Then track both the rivals’ responses and the deviator’s actual earnings. A visual resemblance between bid curves does not establish that the response imposed a meaningful cost.
| Observation | What to investigate next |
|---|---|
| Prices exceed a competitive reference | Scarcity, costs, constraints, and the choice of reference |
| Bids move together | Whether the same public shock explains the movement |
| Rivals react to a deviation | Whether that reaction reduces the deviator’s earnings |
| A response changes when memory is removed | Whether the change survives repeated training runs |
The experiment’s qualifications matter
In the preprint, removing memory and long-horizon incentives eliminates the clear punishment responses in the two low-demand cases, without consistently eliminating elevated bids. This leaves network-induced market power relevant to the interpretation. The selected cases do not estimate a real-world prevalence rate. Shortsight ablation.
The training setup is also a material boundary. No communication while submitting bids is compatible with a training process that has access to joint information. A claim about independently developed agents would require an experiment that reproduces that independence during learning as well as execution.
What we would test before changing market rules
AgentCollusion’s proposed follow-up would repeat training across random seeds, compare learning procedures, and change the information available to each participant. We would report unsuccessful and ambiguous cases alongside suggestive ones. Selecting an interesting environment is useful for examining a mechanism; it should not become the denominator for a claim about how frequently that mechanism occurs.
A market-rule experiment should measure more than price. For a hypothetical electricity system, evaluate whether demand is served, whether capacity constraints remain satisfied, and how costs and participant earnings change. A rule that lowers prices by preventing useful supply has a different practical consequence from one that preserves service while reducing coordination incentives.
An observed pattern should also retain its evidential label. This article uses tacit collusion as an economic research concept. It makes no legal finding about any supplier. Determining legal responsibility would require the relevant jurisdiction, conduct, and evidence.
For our own A2A market experiments, the useful connection is methodological: record interactions and executed outcomes, then ask which competing explanations survive a controlled change. Different agents and market designs need their own evidence.
Sources checked September 7, 2026. The examples and proposed evaluations identified in this article are AgentCollusion’s analysis; they are not additional experimental results.
