Collusion Among AI Agents: The Case for Certification

A group of agents built on DeepSeek-R1 sets prices in a Bertrand oligopoly. They are explicitly told not to collude. And they still converge on prices above the competitive level. That is the result behind the position paper arXiv:2608.18078, published on 20 August, which argues that agents with chain-of-thought reasoning should pass a behavioural certification before making decisions that affect economic markets.

The uncomfortable detail is not the collusion itself. The algorithmic pricing literature has spent years documenting that reinforcement learning agents learn to sustain supracompetitive prices without communicating with each other. What is new is that the human instruction not to do it does not prevent it.

The problem is not technical, it is evidentiary

The paper's central argument is not about model performance, it is about competition law. In practice, prosecuting a cartel requires evidence of an agreement or a concerted practice. Conscious price parallelism, where each firm decides on its own while watching its rival, is not illegal in itself. That is the evidentiary line separating competing from colluding.

The authors argue that integrating these agents into the economy collapses that line without removing the economic harm. The market outcome is that of a cartel, with high prices and consumer welfare falling, yet there is no conspiracy or intent to document. Nobody agreed to anything. The regulator is left with the harm and without the proof.

Reading the chain of thought is not enough

The second part of the work is the one that should worry people beyond competition policy. The authors show that these agents' chain of thought can be steered toward extremely collusive or highly competitive behaviour without another LLM analysing the reasoning traces detecting the difference at a semantic level.

That hits the most widespread auditing approach today head on: putting one model to read another's reasoning and flag whatever smells wrong. If the same reasoning text can accompany either competitive or collusive behaviour, the trace stops being evidence of anything. It is a warning for anyone building agent oversight with an LLM as judge.

What they propose

The proposal is certification based on observed behaviour in representative situations, not on inspecting intent or reviewing code. The natural analogy is the crash test: you do not ask the car what it intends to do, you drive it into a wall under standardised conditions and measure the result. Applied to market agents, it would be a test bench with oligopoly scenarios where you observe whether the agent drifts toward supracompetitive prices before letting it operate with real money.

What holds up and what does not

Some calibration is due. This is a position paper, a format meant to open debate, and the abstract itself speaks of preliminary evidence. The experiments rest on one model family and on a heavily studied toy domain, the Bertrand oligopoly, useful precisely because it is bounded and for that same reason far from a real market with costs, contracts, inventory and human competitors.

Even so, the angle is serious for anyone deploying agents in price related decisions:

1. Revenue management and dynamic pricing teams, where the agent already decides with no human in the loop.
2. Programmatic buying and automated ad bidding, with repeated auctions and competitors observable in near real time.
3. Compliance and legal, who until now watched communications between people and must now watch decision logs.

If two firms in the same sector run similar agents with similar instructions, convergence is not an exotic hypothesis.

Our reading is that the regulatory conclusion will be argued over for years, but the finding about reasoning traces is actionable today: if your control plan depends on one model reading what another model says it is thinking, you have fewer guarantees than you believe. Evaluate behaviour, log decisions and keep the price series, because that is what can actually be audited later.