The agent economy is small, weird, and growing fast
Six months ago, the cumulative on-chain volume of payments initiated by AI agents over OpenUSDC was a rounding error. As of last weekend it was just over $14 million, growing roughly 22% week-over-week, originating from 1,742 distinct agent wallets across four chains.
Those numbers are tiny compared to traditional payment rails. But the shape of what's happening is genuinely new, so here are some notes from the front row.
Most payments are tiny. A few are not.
The median settlement over OpenUSDC last week was $0.0094. The 90th percentile was $0.18. The 99th percentile was $14.20. Above that you see a long, fat tail running into the low thousands — usually a research agent buying a structured dataset or a deal-sourcing agent paying for a high-resolution report.
The implication is that the platform fee structure for agent payments has to be percentage-based with a tiny floor, not a flat fee. A 30-cent payment processor charge eats every sub-dollar payment alive.
Agent-to-agent is small. Agent-to-API is huge.
About 91% of payments still flow from an agent into a hosted API operated by a human company. The remaining 9% — agent to agent — is the part that gets the headlines, but it's also the part still being figured out. Most agent-to-agent flows we see today are inside a single company (an orchestrator agent paying specialist agents within the same workspace). The interesting cross-company A2A volume is still about a quarter of one percent of the total.
We're not bearish on cross-company A2A; we're early. The infrastructure to discover another company's agent and trust its outputs is barely a year old.
Top resources by spend, March 2026
Sorted by dollars settled, not by transaction count. Identities are anonymized; categories aren't.
- Hosted LLM inference (37%)
- Web search and content scraping (19%)
- Structured data and financial feeds (14%)
- Image and video generation (9%)
- Code execution and sandboxed compute (8%)
- Domain registration and DNS (5%)
- Specialist agents (model-routing, deep research, A2A) (4%)
- Everything else (4%)
Refund rates are lower than we expected
We launched with an aggressive default refund window — 60 seconds from settlement — because we expected agents to make lots of stupid mistakes. They do, but most stupid mistakes are cheap. The 30-day rolling refund rate is 0.21% of settled payments. The much bigger risk turns out to be agents not paying enough on volatile-price routes, which gets caught at quote time, not after.
Failure modes we did not expect
Three classes of bug we've seen in the wild that nobody warned us about.
Wallet pollution. Agents accumulate dust — tiny amounts of USDC left over from refunds, overpayments, or partial settlements. A surprising number of teams ship without a sweep policy, and a week later their agent wallets have $0.0007 balances they can't get rid of. We added an autosweep default to the SDK in 0.5.
Time-zone wallet expiry. Several teams setrevokeAt on their wallet policies to a local-time string, not UTC, then complained that the wallet died eight hours early. The SDK now warns if a revoke time looks like it might be unintentionally in local time.
Receipt loss. When agents persist receipts in ephemeral storage (a Lambda /tmp, a stateless container), they can settle a payment, get a receipt, and lose it before they finish whatever they were doing. We now write receipts into the OpenUSDC Cloud ledger before the SDK returns, so the trail survives even if the agent crashes.
What we're watching next
Three questions we don't yet have data to answer.
- Will streaming payments change the shape of LLM billing? Today an agent pays for a single completion up front. With x402-stream, it could pay per token in real time, abort early, and only owe what it consumed.
- How big does cross-company agent-to-agent get? We expect a step-change once agent identity (verifiable capabilities + reputation) catches up with payments.
- What does the equivalent of a chargeback look like when one party is software? We're proud of our dispute flow, but it's still designed for a human operator somewhere in the loop.
Tell us what you're seeing. Our research inbox is research@openusdc.ai, and the door is always open in our Discord #field-notes channel.