AI Agent Payments & Liveness Statistics 2026: Volume, USDC Share, Ticket Size, Working Agents

How to read this page

Three objects get called “the agent economy”: payment receipts, identity registrations, and capital. Different people count them, over different windows, on different chain sets. A registration is not an agent. A receipt is not demand. Float is not agent-held money. Each row states which object it measures; rows we compute ourselves are marked derived. Definitions live in what agentic finance is; the grading scale is the L0–L9 liveness ladder; the rest of the series sits in the research hub.

Payment volume and ticket size

Two readings follow. First, the peak is behind us for now. The largest reported 30-day window cleared ~$24M; the window x402scan showed in September cleared ~$1.45M on a count that stayed in the tens of millions. When value falls and count does not, the ticket collapses — here from ~$0.31 to ~$0.056 — and a ticket collapsing toward dust is the signature of pings and leaderboard farming, not demand. Against that shape, the 25–30% farming estimate reads as conservative.

Second, the cadence is mechanical. Twenty-six million payments across ~21,200 buyers is ~1,230 payments per buyer in thirty days, about 41 a day. No human pays 41 invoices a day. That is the clearest evidence in the data that the buyers are machines, and the reason the x402 brief counts economic transactions — repeat counterparties, value above dust — rather than raw receipts.

USDC share and the stablecoin float

Add the labeled AgentFi figures and you get about $19M. Against a $311B float that is 0.006%; against USDC alone, 0.026%. The gap does not mean agents hold $19M. It means the labeled object is tiny and the unlabeled one is unknown. USDC sitting in an agent wallet at 0% appears in the float and in no AgentFi tile, and no explorer joins identity to wallet to balance, so nobody publishes how much of the float is under machine control. That join is what the graph exists to do. The number it produces — agent treasury utilization, idle against working — is the one to watch, not the float. The parking places are covered in where agents park idle USDC.

The 99% figure is a rail property before it is a preference. It is still a single-issuer dependency: one freeze list, one attestation cycle. We track USDT, PYUSD, FDUSD and EURC share weekly on the data panels; a sustained reading below 97% would be the first real sign of multi-stablecoin routing.

Registrations are not agents

The three counts do not conflict; they are one series. From 220k+ at the start of June to 385,998 in July to 492,792 on 2 September is roughly 2,000–3,000 mints a day over the summer, against a spot rate of 1,560 in the 24 hours before the September snapshot. Velocity is easing. What is counted has not changed: a Registered event, which costs gas and nothing else. Seven validation requests against 125,587 registrations in one window says the part of the standard that checks work is effectively unused. The registries are explained in the ERC-8004 brief.

The liveness funnel

Read the funnel by the size of each drop, because each drop is a different failure. 37% of registrations never set a pointer: abandonment. Four in ten files that parse are not registration files (49.8% valid JSON, 28.3% valid schema): malformed metadata. Of the 17.9% that declare a service, fewer than half answer (6.8%): dead servers. The strict rate is an upper bound on reachability, not a count of economic agents. On our ladder it sits at L5–L6, reach and speak; paid (L7) and funded (L8) come after it, and the default public ranking starts at L7.

Apply the rates and the headline shrinks. 6.8% of Base’s 63,832 registrations is about 4,300 reachable agents (3,500–5,400 inside the interval). The July all-chain rate of 2.24% on the September count gives about 11,000. A defensible sentence: a few thousand to a few tens of thousands of reachable agent identities exist, and a smaller set holds money. Anyone who says 493,000 agents are live is reporting mints. One more caveat: public RPCs silently return empty or truncated log sets for historical queries, so a census that has not verified its own coverage can be wrong in either direction.

Why the sources disagree

• Different objects. Keyrock counts settlements on rails it can see; x402scan counts receipts on one protocol; 8004scan counts mints; Cambrian counts labeled product TVL; Safe counts module transactions. None is a subset of another.
• Different windows. A twelve-month study ($73M), a cumulative-since-launch total (~$50M) and a trailing thirty days (~$1.45M) will never agree, and should not.
• Different chain sets. 27 chains, 29 chains, 22 networks, 15 networks: four views of one standard, four universes.
• Different filters. Farming on or off (25–30% of some windows), spam on or off (empty mints). Most public numbers are off.
• Different tests for “alive”. “Active”, “working service” and “endpoint responds, strict” are three tests with three answers. Only the last publishes its method and a confidence interval.
• Different timestamps and indexers. A $320B high in Q2 and ~$311B on 4 September are both correct. Public RPCs can return HTTP 200 with an empty log set; one public endpoint was short by ~19.6% of agents in a test.

Forecasts, labeled as forecasts

Five orders of magnitude separate the measured base from the forecasts. That is not an argument against the forecasts; it is a statement about what exists today. The line-by-line comparison is in forecasts vs measured reality.

What to watch next quarter

• The x402scan 30-day ticket. Value per payment climbing back toward $0.30 on a steady count means demand; a ticket parked near $0.05 means pings.
• USDC share. A sustained reading below 97% on measured rails is the first sign of multi-stablecoin routing.
• Validation requests. Seven in thirty days is a desert. ERC-8183 job escrow is the candidate that could change it.
• Base strict liveness. The next crawl either leaves the 5.5–8.4% interval or it does not. Movement outside it is signal; movement inside it is noise.
• Labeled AgentFi TVL against float. The ratio is 0.006%. A tenfold move would still round to zero on the float and be a large event for the sector.
• Safe agent share. 37% of transactions with peaks over 75%. A sustained majority would make smart accounts the default agent wallet; the delegation designs are compared in the EIP-7702 brief and the provider comparison. The graph publishes its own L7+ (paid) and L8+ (funded) counts with scrape timestamps; the desk will carry alerts on all of the above.

Sources

• Keyrock, “Who Pays the Agent”, with Coinbase and Tempo data — window May 2025 – Apr 2026, published 2026 — link
• x402scan explorer, trailing 30-day window — observed 2026-09 — link
• Cambrian, The Agentic Finance Landscape Report: Q2 2026 Edition — 2026-06-02 — link
• 8004scan network statistics — snapshot 2026-09-02 — link
• Token Dispatch, “Who Checks the Agents?”, on 8004scan data — July 2026 — no verified link
• maxguryanov/agent-registry-research, Base liveness crawl — through 2026-08-16 — link
• arXiv 2606.26028, Can Trustless Agents Be Trusted? An Empirical Study of the ERC-8004 Decentralized AI Agent Ecosystem — v1 2026-06-24, v2 2026-07-08 — link
• QuickNode ERC-8004 explorer, 30-day window — 2026 — link
• yaojin0609/erc-8004-liveness, indexer honesty notes — 2026 — link
• Agentscan, registration index cited by Cambrian — 2026 — link
• DefiLlama stablecoins dashboard — observed 2026-09-04 — no verified link; the figure is mirrored on our live tape
• IXS vaults, TVL as cited by RWA.io — 2026 — link
• Gnosis Safe agent-share figures — 2026, as reported — no verified link
• Gartner (2028) and McKinsey (2030) forecasts — forecasts, not measurements — no verified link

FAQ