Key Findings
- TRM analyzed approximately $52.7 million across 198.9 million x402 transactions.
- After removing self-payments and anomalous transactions, $25.62 million was identified as likely commerce.
- AI agents represented a minor fraction under TRM’s models, which struggle to differentiate them from automated scripts.
According to TRM Labs, a blockchain intelligence firm, most payment activity on Coinbase’s x402 protocol does not originate from AI agents.
In a report released on Wednesday, TRM evaluated $52.7 million across 198.9 million transactions processed by recognized x402 facilitators on Base, Solana, and Polygon since May 2025. Researchers determined that after filtering out self-payments and other irregularities, only about 0.6% to 7.5% of the remaining commerce appeared to be linked to AI agents.
Myriad: What's next for Bitcoin? Share your prediction.As stated in the report, “The assumption is that a true agent explores across multiple services and products, while an address repeating the same price behaves more like a script hitting one service over and over.” This modeling choice may underestimate the presence of agents, as many could be single-purpose, making repeated payments to one service, which would be interpreted as a script.
Coinbase launched the x402 protocol in 2025 to facilitate payments within web requests. In this process, the buyer receives a quote and authorizes a payment, which a facilitator verifies before completing the blockchain transaction and covering the network fee.
TRM Labs indicated that typical scripts can execute this sequence without the involvement of AI agents. They noted that scheduled jobs, load tests, and self-dealing could yield similar blockchain records, making total transaction amounts inadequate for assessing agent-related commerce.
To refine their analysis, TRM excluded self-payments, bulk transactions from a limited number of payers, and sellers with less than ten buyers, arriving at a figure of $25.62 million in probable commerce. They further filtered for payments broadcast by facilitators, with amounts averaging less than $1. A more stringent test required this payment pattern to persist over several months, alongside public agent registration or payments to multiple sellers. TRM warned that these criteria might overlook genuine agents making repeated purchases from the same service.
In late 2025, the activity encompassed apparent meme-token minting and payments to one AI-analysis service. Early 2026 saw a concentration of volume in a single payment contract before payments for AI services resumed through an agent-payment router around midyear. Throughout the entire period, USDC constituted 99.6% of settled value, totaling $52.47 million.
Companies have been ramping up their investments in agent payments. In August, Binance introduced its Agent OS, which features an x402 payment layer, while Coinbase has targeted agent and payment startups via a $1 million accelerator on its Base platform. In May, Amazon announced AgentCore Payments, developed in collaboration with Coinbase and Stripe, allowing agents to pay for online services using stablecoins.
Beyond assessing usage, TRM also highlighted challenges in attributing responsibility for agent payments.
“On-chain agent registries allow individuals to declare ownership of an agent address,” TRM Labs noted. “However, this declaration is voluntary and not widely adopted among participants.”
TRM advocated for improved registration processes, counterparty reputation information that agents can verify, and monitoring systems designed for handling numerous small payments.
“The infrastructure is already functional. What is needed is accurate registration, counterparty reputation checks an agent can independently verify, and monitoring tailored for volume rather than value,” they stated. “Agentic commerce will require agentic compliance.”