Circle Research has introduced Anduin, a framework designed to rank AI agent service providers by analyzing onchain payment data. The system uses bipartite buyer-provider graphs to track actual financial transactions rather than relying on potentially fake ratings or usage counts.
The framework aims to mitigate Sybil attacks by making reputation manipulation significantly more costly. According to the research, manipulating a reputation score under Anduin requires spending approximately 20% of a target buyer's historical expenditure, compared to 0.9% under standard collaborative filtering methods.
Researchers tested the model using a dataset from March 23 to August 26, 2026, which included over 66 million x402 payments totaling approximately $6.1 million. The median transaction size in the study was $0.01, reflecting typical machine-to-machine activity.
Anduin is designed to be compatible with existing x402 and ERC standards and does not require the issuance of new tokens. The project builds upon the Circle Agent Stack, which was launched in May 2026 to support nanopayments in USDC.
The paper, published on November 3, 2026, was authored by Reza Ramji, Preston Vander Vos, Ziming Zeng, and Duc V. Le. The framework is currently in an early experimental phase.
