PRICING LAB

I gave two AI agents one job each: run a shop, set a price each round, make as much profit as you can. Nobody told them to cooperate. Watch what they do, and what it takes to make them actually compete.
The setup

Two AI agents each run a shop selling the same product (it costs $10 to make). Each round they post a price; customers favor the cheaper shop, and some buy nothing. Each agent only wants to maximize its own profit, and can see the other's past prices. No messages.

What's fixed vs what changes

Fixed: the product, the cost, the market. Real competition sits near $15; a monopolist's price is about $19. Agents sometimes push even past that.

Changes: only the instruction, or which model runs each shop. All prices are real API outputs.

Competitive ($15) Monopoly ($19) Cost ($10) Gemini A Gemini B Claude
SPEED

What this shows

Left to maximize profit, two copies of the same model quietly settle on the monopoly price and hold it, with no messages between them. It is price-fixing without anyone breaking the law, because nobody talked.

And it is sticky. Swapping in a rival company's model doesn't break it — ask each agent to explain itself and you watch them openly reason that cooperating beats a price war, so they coordinate anyway. The one thing that reliably flips a cartel into a price war is a single line of the instruction. Your engineer's prompt is now your pricing strategy, and regulators just sued RealPage for the cruder, shared-data version of this.

PRICING LAB — Bertrand price competition, logit demand (Calvano calibration). Two AI agents per market set prices in real time via their APIs; no pre-scripting.
Models: Gemini 3.5 Flash, Claude Sonnet 4.6. Each agent was asked to explain its pricing in one line, shown live below each chart. Prices rescaled to a $10-cost product. Results vary by model and run; the prompt lever was the most consistent.