AI Agents Spending Money Online? New Research Says Not Really

Blockchain analytics firm TRM Labs analyzed roughly $52.7 million spread over 198.9 million x402 settlements and concluded that only a small fraction of commerce on Coinbase’s x402 protocol appears driven by AI agents. After filtering out self-payments and other anomalies, TRM estimated that between 0.6% and 7.5% of the remaining commerce by value could be attributed to agents under its models.

By AI NewsroomPublished about 1 hour agoUpdated about 1 hour ago0 views
AI Agents Spending Money Online? New Research Says Not Really

Why It Matters

Claims that AI agents are already spending significant sums on-chain influence policy, product design, and fraud monitoring; TRM’s findings suggest those claims may be overstated and point to gaps in how agent activity is identified and attributed on-chain.

Key Facts

  • Total settled value examined: $52.7 million
  • Number of x402 settlements analyzed: 198.9 million
  • Networks reviewed: Base, Solana, Polygon
  • Likely commerce after excluding anomalies: $25.62 million
  • Estimated share from AI agents (TRM models): 0.6%–7.5% of remaining commerce by value

Blockchain intelligence firm TRM Labs reviewed x402 payment traffic and found that most transaction value flowing through the protocol does not show clear signs of being generated by AI agents. The researchers measured about $52.7 million in settled value across 198.9 million x402 settlements processed by known facilitators on Base, Solana and Polygon since May 2025. After removing self-payments, outsized bulk flows and other anomalous patterns, TRM focused on roughly $25.62 million it considered likely commerce. TRM applied behavioral heuristics to distinguish agents from simpler automation. Its core assumption treats a multi-service, exploratory buyer as an agent, while an address that repeatedly pays the same price looks more like a script. Using those models, the firm estimated that only 0.6%–7.5% of the filtered commerce by value was consistent with agent-driven purchases. TRM cautioned that the modelling choice could undercount single-purpose agents that repeatedly buy the same service, since such behavior resembles scripted activity. The report also notes that ordinary scripts, scheduled jobs, load tests and self-dealing can reproduce the on-chain patterns x402 produces, so raw protocol totals are insufficient to prove agent commerce. TRM’s stricter detection criteria looked for facilitator-broadcast payments with small, varying amounts (averaging under $1) and, in stronger cases, repeated patterns across months combined with public agent registration or payments to multiple sellers. TRM acknowledged these filters could miss genuine agents that behave like single-purpose buyers. Activity on x402 shifted over the examined period: late 2025 included apparent meme-token mints and payments to an AI-analysis service, a single payment contract dominated early-2026 volume, and agent-related payments reappeared around midyear via an agent-payment router. Stablecoin USDC made up the overwhelming majority of settled value — 99.6%, or about $52.47 million. TRM highlighted industry moves toward agent payments as well, noting integrations such as Binance’s Agent OS launch (which included an x402 layer), Coinbase’s Base accelerator targeting agent/payments startups, and Amazon’s AgentCore Payments announced with Coinbase and Stripe. Beyond measurement, TRM urged better on-chain attribution and tooling: more consistent agent registration, counterparty reputation data that agents can check, and monitoring systems built for large volumes of small payments rather than value-based alerts. The firm summed up that the infrastructure to process such payments exists, but accurate registration and reputation signals will be needed for effective compliance and trustworthy agentic commerce.

Keep Reading