WallStreetDebug>_ Wall Street, debugged.
Markets
Subscribe
Crypto x AI

AI agents made 137 million crypto payments. Most were not agents paying anyone.

x402, the protocol that lets AI agents pay for data and APIs in stablecoins, is often cited as proof that an agent economy has arrived. A July 2026 study traced every settlement on Base over 280 days. The headline count holds up; the economic activity behind it mostly does not.

136.7Mx402 settlements on Base in 280 days
85%Share of settlements inside operator clustersContext: n=136,708,672 Base settlements · Sep 17, 2025 to Jun 23, 2026 · arXiv:2607.12575 · margin: not disclosed
$187,861Value provably reaching a named service

x402 revives the dormant HTTP 402 'Payment Required' code. An agent asks for a resource, gets a price, signs a USDC authorization, and a 'facilitator' broadcasts the payment on-chain and pays the gas. Coinbase built it; Stripe, Cloudflare, Vercel and Google have integrated it. Dashboards count its settlements in the hundreds of millions.

The study

'How Agentic Is Agentic Commerce?' (arXiv:2607.12575, July 14, 2026), by researchers at City University of Hong Kong, The Chinese University of Hong Kong and Zhejiang University, identifies x402 settlements on Base from their on-chain event and a list of 30 known facilitators, then follows the money through funding and sweep transfers to see where it ends up.

Over the window from September 17, 2025 to June 23, 2026, they count 136,708,672 settlements worth $44,121,383.81, from 491,587 paying addresses to 116,781 recipients. The average settlement is about 32 cents.

Three tiers of activity

The authors sort each settlement by what its on-chain trace can prove. Fictitious payments are a wallet paying itself or money circling in a closed loop: 21.20% of settlements, carrying 54.08% of value Context: n=136,708,672 Base settlements · Sep 17, 2025 to Jun 23, 2026 · arXiv:2607.12575 · margin: not disclosed. Internal settlement is payment between wallets funded by the same operator: 63.78% of settlements and 16.82% of value Context: n=136,708,672 Base settlements · Sep 17, 2025 to Jun 23, 2026 · arXiv:2607.12575 · margin: not disclosed. Only 15.02% of settlements, carrying 29.09% of value, reach a party outside any identified cluster Context: n=136,708,672 Base settlements · Sep 17, 2025 to Jun 23, 2026 · arXiv:2607.12575 · margin: not disclosed.

Together, the first two tiers make up 84.98% of the count, which we round to 85% Context: n=136,708,672 Base settlements · Sep 17, 2025 to Jun 23, 2026 · arXiv:2607.12575 · margin: not disclosed. The authors' robustness check bounds that operator-internal share between 72.38% and 84.98% Context: n=136,708,672 Base settlements · Sep 2025 to Jun 2026 · arXiv:2607.12575, Appendix B · bound as published.

x402 on Base: settlements and value by evidence tier Context: n=136,708,672 Base settlements · Sep 17, 2025 to Jun 23, 2026 · arXiv:2607.12575 · margin: not disclosed
TierShare of settlementsShare of valueValue
Fictitious (self-payment, closed loops)21.20%54.08%$23.86M
Internal settlement (same operator)63.78%16.82%$7.42M
Unattributed (reaches an outside party)15.02%29.09%$12.84M
Total100%100%$44.12M

Who generated the count

  • One campaign in November 2025 involved three wallets that settled 27,089,029 payments to themselves, then stopped within the same two seconds on November 28.
  • One operator hub absorbed 31,551,433 settlements; its funder seeded 337 wallets, and 98.34% of the hub's $415,765 inflow went straight back to that funder Context: n=31,551,433 settlements · Sep 2025 to Jun 2026 · arXiv:2607.12575 · margin: not disclosed.
  • Reproducing all 136.7 million settlements would cost about $355,583 in gas, which facilitators, not the operators, paid.

The supply side

Coinbase's x402 catalog listed 25,163 resources payable on Base. They resolve to only 811 distinct recipients; 624 ever received a payment and 249 earned at least $10 over their lifetime, with median lifetime revenue of $3.96 among the 624. A one-time probe of 910 advertised hosts found 474 still answering with a live payment request.

Other measurements agree

TRM Labs, a blockchain analytics firm, counted about $52.7 million across 198.9 million x402 settlements on Base, Solana and Polygon since May 2025, and estimated that between 0.6% and 7.5% of its screened $25.62 million of commerce looks agentic Context: n=198.9M settlements, $25.62M screened · May 2025 to Sep 9, 2026 · TRM Labs · range as published. In March, an a16z crypto partner contrasted a reported $24 million of monthly agent payments with $3 million visible on-chain and $1.6 million after a wash-trade filter.

Why the count is so easy to inflate

Settlements cost the payer almost nothing, because the facilitator pays gas, and public leaderboards rank services by transaction count. The authors call settlement count a 'Goodhart metric': once rewarded, it measures the incentive to transact rather than adoption. When Coinbase's facilitator began charging a per-settlement fee on January 1, 2026, the daily series recorded its largest sustained fall.

What it shows / What it doesn't

It shows

  • How many x402 agent payments settled on Base over 280 days, and how much of that activity can be traced back to the same operators.
  • A bounded range for genuinely independent agent payments: between $187,861 and $20.26 million.
  • Why transaction counts are a weak measure of AI agent adoption on-chain.

It doesn't show

  • Whether real services were delivered behind internal settlements; the chain cannot show off-chain activity.
  • Which payers are AI agents and which are ordinary scripts. Neither the study nor TRM can tell them apart on-chain.
  • Activity after June 23, 2026, or a full census of Solana.
  • Anything about the value of tokens linked to AI agents.

Why it matters

AI-agent crypto narratives often lean on x402 transaction counts. The best available on-chain measurement says most of that count is operators paying themselves. Agent payments may still grow; the honest current number is small, and traders should discount any pitch built on raw settlement counts.

Primary sources

In our directories

How we used AI: We read the full study and TRM's and a16z crypto's posts, and kept only figures stated in them. Rounding of 84.98% to 85% is ours.

Not financial advice. For information and education only. AI model outputs and research results can be wrong and do not predict future returns. Nothing here is a recommendation to buy or sell any asset. See our disclaimer and methodology.

CryptoAI agentsx402Stablecoins