Blockchain
Circle tests payment-based service rankings for AI agents
On November 13, Circle Research presented a method that uses x402 payment histories to recommend service providers to autonomous agents. The team tested Anduin on 66.1 million settlements and simulated the cost of manipulating its rankings.

Past purchases help agents discover new services
An agent paying for a translation or data access must choose a provider before it knows whether the service meets its needs. A directory can highlight the most-used vendors, but counting transactions or addresses leaves room for fake customers. A service owner can create multiple wallets and circulate its own funds. Circle Research examines what a buyer can infer from its own payment history to make that tactic more expensive.
Anduin starts with the services the buyer has already paid. It identifies other buyers who use the same providers and checks which additional services those customers have paid for. The ranking is specific to each buyer. A provider can rank highly for one customer and remain unknown to another. The study compares three flow algorithms applied to this network. In the scenario tested, transfers among wallets controlled by the attacker are insufficient. To give its candidate weight, the attacker must also spend money with providers the buyer already uses.
A cost difference measured in a simulation
The authors analyzed approximately 66.1 million x402 payments settled on 16 networks from March 23 through August 26, 2026, for a total of $6.1 million. They froze the payment history on June 21 and used the following 30 days of payments to check which new provider each buyer selected. The comparison included 813 buyers who met the study criteria. The Anduin variant called Amplified Flow, weighted by payments and network fees, ranked the provider actually selected first in 22.8% of cases. Collaborative filtering, a baseline that matches buyers with similar habits, reached 23.2%.
The researchers estimated manipulation costs separately. They added a fictitious provider to the graph and calculated the minimum spending needed to rank it first for all 813 buyers tested. Their model included payments to independent providers and a fee of $0.017 per transaction. They found a cost of $8,259 for Amplified Flow and $358 for collaborative filtering. The roughly 23-fold difference applies to this simulated attack and configuration.
New customers have no payment history for the ranking
The study measures the next purchase observed. It does not show whether the service delivered a good result. It also assumes attackers control none of the buyer’s existing providers. A provider that colludes with an attacker or returns the money falls outside the scenario tested. The guarantee therefore depends on the payment relationships the ranking starts from.
A new buyer with no payments has no existing provider for reference. Circle Research points to curated lists of services as a starting point. A new provider also remains unrated until another customer in the buyer’s network pays it. These limitations determine where payment-based rankings can be used. The team presents Anduin as a research project and announced no launch.
