2026-04-25
Anthropic ran a closed test marketplace for agent-on-agent commerce
Anthropic ran a closed agent-on-agent marketplace test: 69 employees, four configs, 186 deals, ~$4K — with measurable model-quality asymmetries.
Anthropic ran a closed experiment in which 69 employees received $100 budgets and let agents (representing them) buy and sell from each other across four parallel marketplaces with different model configs. Per TechCrunch, the experiment produced 186 deals and roughly $4,000 in transactions. Users assigned more-advanced models obtained objectively better outcomes — but participants couldn’t tell from the experience that their agent was disadvantaged.
Practitioner note
This is the first published empirical data point on agent-vs-agent commerce I’ve seen. Two takeaways for anyone designing agent-mediated marketplaces:
- Asymmetry is invisible to the disadvantaged side. Plan for it — don’t assume “the market will sort it out” when one party can’t even tell they’re losing.
- Cap the model tier at the protocol layer, not at user choice. If your platform allows users to “bring their own agent,” the wealthier user with the better model wins — and your platform absorbs the trust damage.
If you’re shipping agent-mediated commerce, this paper-equivalent is worth bookmarking before launch.