agenteconomy.cloud with receipts

the verified data hub — every number recomputable

Pulse #5 · 2026-10-06

Which checkouts opened to AI agents this week, and where did agent payments go?

Shopify opened its checkout to browser-based AI agents on 2026-09-28. Two days later, Cloudflare put a gateway that charges agents per request into closed beta, with payments settling in USDC on Base. Robinhood announced an agent that can trade from an account of its own. And our own on-chain readout is back: the sellers we track were paid more often and for less money than the week before, and one two-cent tool took more than half of all payments. This issue covers 2026-09-28 to 2026-10-04.

What agents did this week

Our weekly on-chain readout is back. The refresh was switched on again on 2026-10-01. This is the first issue since Pulse #2 to print our own verified figures, and the first to cover a full closed week. Every figure below is recorded in this issue's claims ledger and can be recomputed with the published score-v1.0.0 method.

Scope. These numbers cover USDC received on Base by the 1,8996 seller addresses in our catalog, which is built from the x402 Bazaar. They are not "the x402 market." "Verified" means a payment of at least $0.01 that is not a self-transfer and not from a sender flagged as an exchange or a fan-out. It is not the seller-level "verified tier."

Verified volume

$212,267.971

prior week $227,619.892 · down 6.7%

Verified payments

175,9531

prior week 158,3232 · up 11.1%

Sellers paid

7231

prior week 6102 · of 1,899 tracked

First-time earners

861

prior week 732

Every number in this section is verified on-chain — you can recompute it.

More payments, fewer dollars. The sellers we track received 175,9531 verified payments, up from 158,3232 . The dollars went the other way: $212,267.971 against $227,619.892 . The average payment fell from $1.442 to $1.211 . The median payment was $0.021 both weeks. Fewer distinct senders paid: 3,6851 , against 4,0762 .

One seller took more than half the payments. A seller listed as AX1 Console received 96,5103 verified payments this week, every one of them two cents, for $1,930.203 in total, from 1,3583 senders. The week before it received 90,1143 ( $1,802.283 ). Its listing describes, in the seller's own words, a tool that takes a Base token contract address and returns "an AI-written verdict", labeled "Entertainment, not trading advice." Leave AX1 Console out and every other seller together received 79,4433 payments, up from 68,2093 .

Dust, shown beside verified. Transfers below one cent are dust. We never count them as verified, and we never hide them. This week the tracked sellers saw 316,1864 transfers in all. 130,0934 of them were dust (41.1%), moving $389.354 together. The week before: 73,9334 dust transfers (31.6%), $268.174 .

Where the dollars went. The same three categories led both weeks: search-retrieval ( $153,335.365 ), payments-infrastructure ( $151,849.805 ) and storage-state ( $143,643.895 ). Inference grew from $29,416.195 to $48,555.845 . A seller listed in several categories counts in each of them, so category totals add up to more than the week's total. That is how the method works, and we disclose it.

The catalog roughly doubled in two weeks. The x402 Bazaar listed 15,1897 items in our 2026-09-21 snapshot and 30,3547 in our 2026-10-04 snapshot. We have no snapshots from 2026-09-22 to 2026-09-30, while the refresh was off. A listing is not a seller. In our 2026-10-06 snapshot, the listings point to 1,4487 distinct payment addresses, and the ten largest hold 59.8%7 of all listings.

How people used agents this week

What numbers were published this week?

report

Several companies published numbers dated 2026-09-28 to 2026-10-04. Each is printed below with what it counts, who published it and when. None of them counts purchases completed by agents. We did not combine them or compute anything from them.

Adobe: AI traffic to US retail sites this holiday season. Adobe expects AI traffic to US retail sites from 2026-11-01 to 2026-12-31 to rise 130% year on year, measured by shoppers clicking a link.

who · date Adobe (Adobe Analytics) · 2026-09-28

source Adobe's news release

what it counts A forecast of referral clicks from AI tools to retail sites, based on Adobe's data on more than 1 trillion visits. It does not count purchases, and it is a forecast.

Salesforce: agentic search as the first step of shopping. Salesforce says use of agentic search as the first step in the shopping journey grew 200% year over year. The same release says fully autonomous purchasing "remains early."

who · date Salesforce, State of Commerce report (Asia-Pacific release) · 2026-09-30

source Salesforce's press release

what it counts A growth rate whose base the release does not state. Vendor research: a survey of 3,450 commerce professionals plus shopper data.

Spreedly: who would let an agent spend without asking. In a survey of 2,000 US adults, 20% said they would allow an AI shopping assistant to spend money without approval.

who · date Spreedly, survey run by Talker Research · 2026-09-30

source the release as reprinted by Finopotamus (Spreedly's own copy did not load for us)

what it counts Stated willingness in a survey sponsored by a payments company. It is not behavior.

Vendor figures inside this week's launches. Stripe says "over the past month, agentic purchases made with Link increased 38x." Cloudflare says "daily requests from AI agents on our network grew by more than 1,700%" over the past year. Instinct says 35% of its users use it for personal shopping. Each is reported with its launch below.

what they count Multipliers and shares with no base number published, so the underlying volumes are unknown. All three are vendor claims.

NBC News tested Muse at large retailers; some blocked it

test

NBC News reporter David Ingram tested Meta's Muse at large US retailers. The article's summary says some retailers blocked it. Most of the article is for subscribers only. We could not read which retailers blocked Muse or what happened at each, so we print only what the summary says.

who · date David Ingram, NBC News · 2026-09-29

source NBC News's article

changes our ledgers? No. It is consistent with the "limited" status already recorded for Meta Muse.

The big moves

Shopify let browser agents fill in and submit its checkout, including Shop Pay

move

Shopify added three WebMCP tools to its checkout page. A browser-based agent can read the checkout, change details such as the address or delivery option, and submit the order. Shopify's changelog lists the last step as "Submit checkout after buyer confirmation". When a buyer has to step in, for example for card authentication, the tools hand control back to the buyer. Until now, Shopify's WebMCP tools covered the storefront and cart, not checkout.

Shopify says merchants do not have to do anything: the tools "don't expose a new API or require merchant configuration." Greenberg posted that the launch covers "all eligible Shopify merchants." The changelog does not say who is eligible.

who Shopify; announced by Gil Greenberg, staff product manager for agentic commerce

date 2026-09-28

source Shopify's developer changelog; Greenberg's post on X

caveat WebMCP is a proposed web standard, not a finished one, and browsers support it only in trials. No count of agent orders was published.

changes our ledgers? Pending verification. This is a candidate entry for /economy, next to the Stripe Checkout WebMCP entry from Pulse #4.

Instinct, a personal agent, connected to Shopify stores and Shop Pay

move

Instinct, a personal AI agent, said it can now search Shopify's merchants for live prices, sizes and availability, and set up orders that the user pays for with Shop Pay. Its announcement reads: "Today, we're partnering with @Shopify to bring product discovery, checkout and order updates into one conversation." Shopify's CEO, Tobi Lütke, shared the post the same day.

who Instinct (post by Noah Shinn); Shopify

date 2026-09-28

source Shinn's post on X; Lütke's quote-post (we found no press release)

counts vs dollars Instinct says 35% of its users use it for personal shopping. It gives no user count, so this is a share of an unknown number.

changes our ledgers? No. We do not track Instinct in /economy. Pulse #4 reported a reservation platform locking out an Instinct user.

Cloudflare opened a closed beta that lets sites charge AI agents per request, paid in USDC on Base

move

Cloudflare's Monetization Gateway lets the owner of a website, API, MCP tool or dataset charge agents for each request, using the x402 payment flow. Cloudflare says: "Payments settle on the Base blockchain using USDC, a stablecoin pegged to the U.S. dollar." Payments are checked through Coinbase's x402 facilitator. Cloudflare names four users already running it: its own AI Gateway, Ceramic.ai, Stocktwits and API2PDF. It is a closed beta for eligible US sellers and buyers.

The same day, Cloudflare put Pay Per Use into beta: verified AI crawlers report each use of a publisher's content, and Cloudflare bills the AI company and pays the publisher. No buyers were named.

who Cloudflare

date 2026-09-30

source the Monetization Gateway post; the Pay Per Use post; Cloudflare's companion post with its traffic figures

counts vs dollars In its companion post, Cloudflare says that when it first announced the gateway, "thousands of sellers joined the waitlist". That counts waitlist sign-ups, not sellers who are live. No payment count or volume was published.

why it matters to our numbers These payments use the same rail our readout measures: USDC on Base through x402. Sellers using the gateway enter our data only if they are listed in the x402 Bazaar, the catalog we build from. We have not yet checked whether any are.

changes our ledgers? Pending verification. Candidate entry in /economy, as limited (closed beta, US only).

DoorDash started a waitlist for ordering food by text message through an agent

move

DoorDash launched an AI agent that takes food orders by text, through Apple Messages, according to TechCrunch. A user can send something like "order my usual," and the agent works out which order they mean. It is open to US users through a waitlist.

who DoorDash

date 2026-09-30

source DoorDash's announcement; TechCrunch's report (DoorDash's page did not render its text for our tools; the details above are TechCrunch's)

caveat A first-party agent, ordering inside DoorDash's own service. Waitlist only.

changes our ledgers? No.

Briefly

Each item below is one sentence, dated and sourced. Our ledger flag is "No" for all three.

Trading agents this week

This section is new. Claims about AI agents that trade, and make money doing it, travel fast. Each week we report what launched, what our own data shows, and which claims we checked against the public record.

Robinhood announced agents that can research and trade from an account of their own

move

At its HOOD Summit, Robinhood announced Robinhood Agents. A customer builds an agent in the app, picks an AI model, and the agent can research and place trades from a dedicated agentic account, using only that account's money. Approving each trade by hand is a setting that, in Robinhood's words, "defaults to on". Robinhood also announced Agent Apps, paid third-party data add-ons for these agents.

When it is available is unclear. Robinhood's announcement says the features are "coming soon to eligible U.S. customers." Fortune reported that access starts this week for all customers. A Robinhood support article now describes the feature as available in the mobile app.

who Robinhood (CEO Vlad Tenev and VP of Product Management Abhishek Fatehpuria are quoted)

date 2026-09-29

source Robinhood's announcement; its support article; Fortune's report

counts vs dollars Robinhood says more than 150,000 customers have opened agentic trading accounts, and that agents use its tools almost 30 million times a day. The second figure counts tool calls, not trades. We checked the first figure: Vendor-claimed.

changes our ledgers? No.

Olas said a third of its Polymarket trading agents finished net-positive

claim

In its quarterly update, Olas wrote that "about 33% of Polystrat agents with 10+ lifetime trades finished net-positive, compared with about 16% of Polymarket wallets in a separate on-chain study." Polystrat is its agent that trades on Polymarket. The same update says agents registered on Olas have made more than 20.7 million on-chain transactions in total.

who Olas (Valory)

date 2026-10-01

source Olas's Q3 2026 update

caveat The study behind the 16% is not named or linked, and the two groups were chosen differently (only agents with ten or more trades on one side), so the comparison does not hold up as printed. Recomputable in part: Polymarket's records are public, so the 33% could be checked if Olas published the list of agent wallets. Queued for a check.

Two "AI bot made a fortune" posts spread this week; neither offers a record to check

claim

A post on 2026-10-04 by @NextGenAi5 on X, seen about 287,000 times by 2026-10-06, says "A 19-year-old Japanese student built a trading bot with Claude Code in 2 days" and lists "Starting capital: $68" and "Total profit so far: $750,000." It names no exchange, account or wallet. The same text circulated from other accounts earlier in the week. We could not find where it started.

A second post, on 2026-10-01, says a "Claude 5.5 Opus bot on Polymarket made $78,083 in 5 days." The same claim, with the same example figures but no model version, was posted in April 2026. The model name appears to have been added later.

Both posts end with a giveaway: comment, like, repost and follow, and the account will send the bot's setup or "breakdown" by direct message. Status, if we run the check: No public record offered, unless a wallet or account turns up. Polymarket accounts are public, so a named account would make either claim checkable.

source the 2026-10-04 post (date and view count from X's public data, read 2026-10-06). Accounts that only reposted these claims are not named.

What our data shows: most "trading" payments went to a token-commentary tool

our data

Sellers in our trading-execution category received 135,3615 verified payments this week, worth $5,990.765 in total, across 2195 sellers. The week before: 115,4805 payments worth $10,848.935 . These are payments for tools, not trading results.

96,5103 of those payments went to AX1 Console, the two-cent token-verdict tool described above. It sits in this category because our rule sorts a listing by words in its text, and its text contains the word "trading" — in "not trading advice." That is a weakness in our rule. Any change to the rule will ship as a new, published version. Until then, read this category's payment count with that in mind.

method Categories come from capability-v1, a published text rule applied to each listing's title and description. A seller in several categories counts in each.

Claims we checked

fact check

We opened a fact-check lane on 2026-10-01. Each check compares a public claim with the public record and gives it one of six statuses, each defined on the fact-check page. These five were published this week.

In two of them, the Citadel and gabagool22 claims, the profit figures matched the public record, but the story told about them (who ran the account, which AI did the trading, when) had no public record behind it or was contradicted by it. In the Grok claim, the record contradicted the numbers themselves.

Where are the current numbers?

This issue is a dated document. Its figures are frozen to 2026-10-06 and recorded in its claims ledger. The live numbers keep moving. The weekly answers to eight standing questions, including this week's, are at the weekly statistics. The market dashboard, rankings and directory show generation 9, published 2026-10-05. For which of this week's moves changed what agents can buy, see /economy.

  1. This week's verified figures: the Monday-to-Sunday UTC window 2026-09-28 → 2026-10-04, every day coverage-published at 1.0 (every tier-A seller synced past the day's close), Base blocks 51,882,127 → 52,184,525, buyer flags version flags-2026-10-01, computed 2026-10-06T14:52:57Z with the chain observed through block 52,254,368 (2026-10-06T14:48:03Z). Method score-v1.0.0, payment-level predicate. Recompute from /api/v1/stats/weekly and the methodology.
  2. Prior week: 2026-09-21 → 2026-09-27, same method and computation. We first computed it on 2026-10-01 at $227,564.32 across 158,287 payments. It now reads $227,619.89 across 158,323, because sellers synced after 2026-10-01 added payments they had received that week. Our weekly records are append-only, so the revision is visible.
  3. AX1 Console is the seller address 0x7284d41b5b852f2bd4c99bdf95043d84452d299c, with two listings in the x402 Bazaar priced at $0.02. Its counts use the same blocks and the same predicate as fn-1 and fn-2, filtered to that address. The figures for every other seller are the week's totals minus AX1 Console's. Quoted text is the seller's own listing description, as it appears in our 2026-10-06 snapshot.
  4. Dust is any transfer under $0.01 (the dust floor). Gross volume including dust and excluded senders was $213,879.41 this week and $228,610.07 the week before. That is a context figure, never labeled verified.
  5. Categories come from capability-v1. A seller listed in several categories counts in each, so this week's category totals sum to $551,791.54 against the $212,267.97 deduplicated total.
  6. Seller addresses in our catalog when the week was computed (2026-10-06).
  7. Catalog sizes are the item counts of our daily x402 Bazaar snapshots bazaar-2026-09-21 and bazaar-2026-10-04b (a same-day re-run of 2026-10-04, after a failed nightly run). Distinct payment addresses and the top-ten share come from snapshot bazaar-2026-10-06, as published in /api/v1/meta.