The most concrete evidence yet that machine-to-machine payments are happening comes not from a roadmap but from transaction flows: AI agents are now buying data, compute power and online services on their own, and the settlement asset is overwhelmingly a stablecoin. Reports from Blockonomi and TokenPost this week describe agents executing independent transactions, predominantly in USDC, as payment companies compete to build the rails that let software hold and spend money. The pattern matters because it is the first class of crypto usage where the end user is not a person at all — and stablecoins, not bank rails or cards, are what the infrastructure has converged on.
What the agents are actually buying
The current spend profile is narrow but real: data feeds, API access, compute capacity and subscription-style online services. These are exactly the purchases that suit autonomous software — small, recurring, metered, and priced per call. An agent that needs a web-scraping API or a vector database does not need a checkout page; it needs a programmable balance and a payment primitive it can call like any other function. That is why stablecoins dominate: USDC on cheap, fast chains behaves like a programmable dollar that can be granted, spent and audited entirely in software. The blockonomi report notes these transactions are predominantly USDC-based, which is consistent with what the infrastructure stack — Coinbase Agent Payments, Skyfire, x402-flavored flows — has been built around.
Why stablecoins took the early lead
TokenPost frames the competitive dynamic plainly: payment companies are racing to build infrastructure for software as the next major class of crypto users. Stablecoins won the early lead for structural reasons. Card rails require human authentication patterns agents cannot satisfy; bank transfers are slow and jurisdiction-bound; and native crypto assets introduce price risk that a metering agent cannot budget around. A dollar-pegged, chain-native asset gives agents a stable unit of account with 24/7 settlement. The consequence is that the agent payments market, however small today, is defaulting to stablecoin rails before the traditional payments industry has shipped a comparable product. Whether Visa and Stripe can insert themselves at the API layer — rather than the settlement layer — is now the open competitive question.
The volume caveat
Enthusiasm about agents as a user class should be tempered by the actual numbers. Agent payment volumes remain depressed relative to the infrastructure built on top of them; earlier reporting showed flows like x402 down sharply year-to-date even as Coinbase, Stripe and exchanges keep shipping agent tooling. The current reports describe a market that is “just getting started” — an accurate framing, but one that describes pilots and early integrations rather than a settled economy. If agent commerce follows the pattern of early internet payments, the first use cases (data and compute) may stay small even as the tooling standardizes, with the larger spend arriving only when agents routinely act on behalf of consumers and businesses.
What to watch
Three indicators will tell you whether this thesis is compounding. First, whether non-USDC settlement share grows — USDT and other stablecoins have agent-facing integrations, and concentration in one issuer is a systemic point of failure. Second, whether spend diversifies beyond data and compute into services with larger ticket sizes, which is where traditional payment processors would feel real pressure. Third, whether reconciliation, spend limits and audit tooling mature — the plumbing that finance teams require before any corporate agent is allowed near a real budget. The infrastructure race is well ahead of the demand curve; the next data points will show whether the curve bends up to meet it.