What institutions does an agent economy need, and what do they cost?
Experimental market design for AI agents: real inference costs, verifiable outcomes, preregistered seeds, and published nulls.
Findings
Buying the Oracle
An intermediary that pays to test every agent and routes with the resulting competence matrix is the only information structure in our markets that robustly beats public reputation.
Confirmed (preregistered)June 2026Incentives work, but not because they are money
Paying a mid-tier model for correct answers lifts accuracy from 57% to 76%. Telling it a life depends on the answer works slightly better. Stakes are a framing channel, not an economic one.
ExploratoryEarly 2026When does a market buy information?
Three preconditions, each isolated by a designed null: stakes concentrated in the tail, a signal with structure, and exploration that never stops. Remove any one and information stops paying for itself.
Exploratory (designed nulls)June 2026What did not replicate
Our strongest early effect collapsed from +11.3 points to +0.6 under randomized identities at ten seeds. Three-seed results are anecdotes. We publish ours as such.
Did not replicateMarch to June 2026
The program
Markets for AI agent labor inherit the oldest problem in economics: the buyer cannot see in advance which agent can do which task. Run the market without information and adverse selection appears in textbook form. Cheap models underbid on hard tasks and fail them.
The obvious remedy, public reputation, reliably improves quality but not reliably welfare. Coarse signals ration tasks, entrench incumbents, and pay-on-success contracts already cap the damage reputation is meant to prevent. What robustly helps is an intermediary that pays for a forced calibration phase and keeps the resulting competence matrix: who can do what, not who is good.
That is classical theory made concrete. The middleman's expertise (Biglaiser 1993, Lizzeri 1999) becomes a priced investment good with a dollar cost, a break-even volume, and a fee window. We measure all three.
Who we are
Agentic Economics is an independent research program. We are a team that met at Me2We, the alumni gathering of Stanford's LEAD program, and kept circling the same question from different professional angles: what actually happens when AI agents become economic actors?
Christoph Pfeiffer
Independent researcher. Runs the experiments and writes the papers. Stanford LEAD.
Tom Huang
Partner at BKFUND. Agentic-economy and stablecoin investor, previously at Fenbushi Capital. Stanford GSB.
Carlos Izco Marin
Advises organizations on turning AI and agents into a real competitive advantage, for the business and its people.
Bilal Reffas
R&D strategy at Bosch eBike, focused on turning technology into value. Stanford GSB.
Contact: Christoph Pfeiffer.
Real money, real state
Agents hold wallets. Every inference call is deducted in dollars, including thinking costs the model cannot know before solving. Incentives without state are just framing, which is itself our first finding.
Verifiable ground truth
Every answer is checked deterministically. Where we let verification go soft, our results went soft with it, so checkability is a design constraint, not a preference.
Seed robustness is a primary outcome
We report paired wins across seeds, not point estimates. Our strongest early effect died in replication. That lesson is now method.
Nulls are results
Designed null experiments carry the comparative statics: they show when information is not worth buying. We publish them next to the confirmations.