Method

Agents in our markets hold wallets. Every inference call is deducted in dollars, including thinking costs that scale with difficulty and are unknown to the model before it solves. Bid costs are booked too: selection effort is explicit, not free.

This constraint exists because of our first result. Prompt-level incentives lift performance, but so does any high-stakes framing; money is not a special channel at the prompt level. Testing genuinely economic mechanisms requires state: budgets, costs, consequences.

Every answer in the core experiments is checked deterministically. No LLM judges on the critical path.

We know what happens otherwise, because we tried it: in an early embedding-reputation experiment scored by a soft LLM judge, market and central planning were indistinguishable. Where verification goes soft, results go soft with it. Checkability is a design constraint and, we suspect, a constraint on where agent markets can work at all.

LLM market experiments carry enormous seed variance. Our strongest early effect, +11.3 points from public reputation, collapsed to +0.6 under randomized identities at ten seeds. A later 3-of-3-seeds welfare effect collapsed at five.

We therefore report paired wins across seeds next to means, and we treat point estimates from a handful of seeds as anecdotes, including our own.

Exploratory results are labeled as such. A claim is framed as confirmed only after a run on fresh, pre-declared seeds against a success criterion stated in advance. The intermediary result is the first in the program to pass this bar: the criterion was at least 7 of 10 paired wins, the run delivered 8 of 10, and the effect size shrank from the exploratory +22% to +14.5%. We report the smaller number.

We also document where we walked into the garden of forking paths ourselves: the exploratory seed sets were extended sequentially while results were visible. That is exactly why the confirmatory protocol exists.

Experiments designed to remove one precondition at a time, and showing that the effect disappears, carry the comparative statics of the theory. We publish them next to the confirmations, with the same prominence.

We claim nothing about the models' intentionality or strategic cognition. LLM agents behave as if they respond to incentives: reproducible input-output patterns, presumably artifacts of training data. For mechanism design, that is the relevant fact regardless of its origin, and every claim on this site should be read inside this frame, in this setup.