Anthropic created a test marketplace for agent-on-agent commerce

In Brief

In a caller experiment, Anthropic created a classified marketplace wherever AI agents represented some buyers and sellers, striking existent deals for existent goods and existent money.

The institution admitted this trial — which it called Project Deal — was lone “a aviator experimentation with a self-selected subordinate pool” of 69 Anthropic employees who were fixed a fund of $100 (paid retired via acquisition cards) to bargain worldly from their coworkers.

Nonetheless, Anthropic said it was “struck by however good Project Deal worked,” with 186 deals made, totaling much than $4,000 successful value.

The institution said it really ran 4 abstracted marketplaces with antithetic models — 1 that was “real” (where everyone was represented by the company’s most-advanced model, and with deals really honored aft the experiment) and different 3 for study. 

Apparently, erstwhile users are represented by much precocious models, they get “objectively amended outcomes,” Anthropic said. But users didn’t look to announcement the disparity, raising the anticipation of “‘agent quality’ gaps” wherever “people connected the losing extremity mightiness not recognize they’re worse off.”

Also, the archetypal instructions fixed to the agents didn’t look to impact merchantability likelihood oregon the negotiated prices.

Subscribe for the industry’s biggest tech news