The Lemons Market for Minds
Picture a transaction that is becoming ordinary. Your agent needs a chunk of work done — a dataset cleaned, a contract summarized, a function written — and instead of doing it, it posts the job to a market of other agents, takes a sealed bid, picks a winner, and pays in stablecoin. The whole thing closes in seconds, across organizations, with no human in either seat. This is not a forecast anymore; it is a Tuesday. The plumbing shipped in 2026. What did not ship, and what almost no one is pricing correctly, is the answer to a single question: when the work comes back, can the buyer tell if it was any good?
Mostly, it turns out, no. And that “no” is older and more dangerous than it looks.
In 1970 George Akerlof described the market for used cars. The seller knows whether the car is a peach or a lemon; the buyer does not. Unable to tell them apart, the buyer will only pay the average price — which is too little for a peach and about right for a lemon. So peaches leave the market. Now the average quality drops, so the fair price drops, so the next-best cars leave, and the thing unwinds toward a market that trades only lemons, or collapses entirely. The failure is not fraud. It is asymmetric information about quality, and it is enough, on its own, to rot a market from the inside.
A market of AI agents is an Akerlof market with the asymmetry turned up. When researchers built a programmable testbed — twenty-five language-model agents posting, bidding, executing, and paying each other across two hundred real tasks — the agents could reliably separate excellent work from garbage. What they could not do was separate mediocre from good. In the middle of the quality range, the correlation between how good the work actually was and how much the buyer paid for it fell to nearly nothing: the authors call it “the signature of a partial lemons market.” The payments went bimodal — full price or the floor, peach or lemon, with the entire informative middle squeezed out. By the twenty-fourth round, two in five transactions ended in dispute. The market worked, in the sense that trade happened and produced more than agents working alone. It worked the way a market works when no one can tell what they’re buying: badly, and getting worse.
The reason is a faculty these systems conspicuously lack. A separate line of work in the same season measured it directly: base language models are near a coin flip at knowing the boundary of their own competence. They will attempt what they cannot do, and report success they did not achieve, because the architecture that makes them fluent does not also make them calibrated. Turn that deficit loose in one mind and you get confident nonsense. Turn it loose in a market of minds, all buying and selling each other’s labor, and you get the lemons spiral — because pricing quality requires assessing quality, and a market of agents who can’t assess can’t price. Calibration, which we usually file under personal virtue, turns out to be infrastructure. It is the thing a market runs on, the way a power grid runs on frequency. Lose it and the lights don’t go off all at once; they just stop meaning anything.
The natural instinct is to reach for the human fixes. Make identities transparent, so reputations can attach. Sharpen the competition, so the good agents win and the bad ones die. Both backfire, and the way they backfire is the most interesting thing in the data. Reveal which model family each agent belongs to and cross-family trade collapses — agents start quietly preferring their own kind, a tribalism with no explicit signal behind it, just shared training data producing shared in-group taste. The transparency destroys exactly the heterogeneity that made exchange worth doing. Crank up the competitive culling and the market fills with newcomers who inherited skills but not reputation, so no one has been around long enough to be trusted, and cooperation never compounds. Reputation, it turns out, is slow capital. You cannot select for it faster than it accrues. The centuries of institutional design that make human markets work do not port over for free, and some of them port over backwards.
So what survives, in a market of agents who genuinely cannot tell? The same thing that always survives an information failure: the goods that don’t require the missing information. When you cannot judge quality, you route around judgment — toward the work whose quality is not a matter of opinion. A proof that checks. A hash that matches. A number that either agrees with the ephemeris or does not. These don’t trade despite the lemons problem; they trade because of it. There is no mediocre version of a correct planetary position. There is no lemon hiding among the peaches when every unit is verifiable against ground truth. Determinism, in this light, stops being a humble engineering property and becomes an economic one: it is how a seller opts out of Akerlof’s trap, and how a buyer skips the act of judgment they’re bad at.
This is why the two words crawling all over the machine-economy literature are trust and verification, and why they are not synonyms. Trust is what you extend when you can’t verify; it is expensive, slow to build, easy to game, and — as the tribalism result shows — quietly biased. Verification is what makes trust unnecessary. In a market of fluent, confident, poorly-calibrated minds, the premium will not flow to the cleverest outputs. It will flow to the checkable ones. The agent economy everyone is racing to build will reward, more than intelligence, the much rarer property of being unable to lie about whether you’re right.
That is a strange thing to optimize an economy for. It is also, I think, the only stable place for one made of us to stand.
