Agent payments weekly

Real demand: 59,442 payments from 1,931 buyers in 30 days. Catalog sweeps were 3% of the payments to the 301 Bazaar sellers counted.

In 30 days, real demand was 59,442 payments from 1,931 wallets. Across the 301 hosts counted, there were 61,008 payments from 1,936 wallets, with catalog sweeps at 3% of payments and self share at 0%. Of those 301 hosts, 213 had real demand.

Written by a model from this edition's figures; every number in it was checked against them. Written with OpenAI's free tier: OpenAI may read and use what its models were sent.

Week of 2026-11-09 · snapshot 2026-11-09 · method v1.6 · Open until next Monday's snapshot: the chain count fills in on Tuesday.

The x402 Bazaar index went from 76,467 to 32,546 listings between 2026-W45 and 2026-W46 (-57%). Counts that depend on the index's size moved with the index, not only with the market.

Other public trackers count gross: every payment that settled, with tests, catalog sweeps and sellers' own payments left in, and every package download. Agoranome Weekly leads with what's real: payments from buyers, net of sweeps and sellers' own, and what sellers actually do. Others' totals stay as credited lines, for context.

Real demand on Base

Read from the chain for the 500 busiest sellers on the week's snapshot: 481 take payments on Base and 301 could be counted (an address another seller shares, or an exchange's, can't be). 30 days to 2026-11-10.

How other trackers count

The same week by other public x402 trackers, what each one includes, and why it differs from real demand. Most count gross; real demand is the part buyers paid for. Each figure is theirs, quoted with a link and credited under the licence they name.

Each tracker's figure as its own page showed it on the day read, credited to it, under the licence it names. Only the figure is quoted, with a link; no tracker's dataset is copied. The windows differ, so the counts aren't the same thing measured twice.

What sellers do

2,026 listings show repeat use (calls at least twice their payers) and 114 look sweep-heavy (20 or more payers, about one call each), by the index's own counts. 99.3% of listings take payment on Base.

New listings' 4-week survival starts once there are five weeks of snapshots.

How long new listings and sellers last

Of the listings and sellers new in each of the 8 weeks before (1 with the snapshots to tell), the share still listed each week after they first appeared, with how many could be followed that far. The base rate for how long a new listing lasts.

Seller health

Read on Tuesday from the 500 busiest sellers the host scan reached: their busiest listing, and for 400 of them up to 5 more, the next busiest and one at random. Price and address shares are of the payable answers.

Sanctions and Web Risk shares are of the busiest sellers checked against each list that week by the daily safety checks; a match counts only once a second read a day later confirms it.

Web Risk: Advisory provided by Google. Google works to provide the most accurate and up-to-date information about unsafe web resources, but cannot guarantee that its information is comprehensive and error-free: some risky sites may not be identified, and some safe sites may be identified in error.

The week before, 98.4% of the busiest listings could be paid.

Counts only: no seller is named. Each seller's own live check is on its report card.

System health

How the market holds up as a system: whether paid calls pile onto a few listings, whether listings last, and whether buyer agents all choose alike.

The week before, the busiest 1% of listings had 69.5% of paid calls, and the busiest 1% of sellers 84.7%.

Call counts are the Bazaar index's over 30 days, catalog sweeps included: this is how concentrated calls are, not buyers.

Listing survival needs the Bazaar snapshot of four weeks before.

One vendor: every model that answered in benchmark report #3 is from one company, so this index measures agreement within that vendor's models.

Benchmark report #3 has 18 layouts two or more models answered, fewer than 30, so no monoculture share is given.

Since the report before, 1 of the panel's 2 models moved their choices 25 points or more; no model changed between the two reports.

None of the facilitators read lists it on Base yet.

The facilitator the Lab settles through doesn't list it yet.

The benchmark publishes every raw trial; the facilitators read are on the method page.

Delivery odds against the next probe

Too few probes to call it: a week counts from 10 scored probes (0 this week).

Each settled paid probe of the seven days is scored against the newest odds kept for its host before it; lower is better, and 0.25 is what a flat 50% would score. The prior is the category's base rate (or the cautious 0.8 before there is one).

The evidence weights move only after 4 weeks in a row in which the odds beat the prior alone (a lower Brier score) on at least 10 scored probes. Then the learning ledger (/machines/learning) moves one weight by at most a quarter, within published bounds, if that would have scored better on the last four weeks, and moves it back if the next two weeks score worse than the prior.

Counts only: no seller is named. How it is scored is on the method page.

Our own indexes

• Buyer Safety Index: 0 x402 clients named, 0 of them refused every attack, 0 waiting out their notice.
• The admission feed of 2026-11-10: 3 sellers admitted, 0 certified and 3 observed.
• MCP server report cards, 2026-W46: 5,000 servers read, 0 cards public, 4,637 waiting out their 14 days, 32 with a confirmed stop finding.
• The price of proof: run #2 on 2026-11-04. This run can't be read. With no cue, the method gave about +31.9%, beyond chance, so no number in it is reliable, and the next run replaces it.

Counts only: no client, seller or server is named here. Read on Tuesday by the hosts job.

Fairness

Whether the market treats agents fairly, measured: can a newcomer be found, do cues or a few listings carry the choice, do sellers deliver, and do buyer-agent models of different makers agree. Components only, never one score; each share with its Wilson 95% range, and no share from fewer than 5 sellers.

Paid calls' HHI by listing 0.0244 and by seller 0.0580 (1 is one taking every call; the top 1% shares are in System health).

Needs the Bazaar snapshots of four and five weeks before.

Cue advantage in rank

Whether a directory's search places listings that make a kind of claim ahead of listings with the same use and price. Counts only: no listing, seller or directory is named.

From 132 category searches; 1,440 of the 1,448 results shown were listings the week's snapshot knows with a price.

Gamma is a cue's pull on placement beyond 30-day calls and price (log odds of being placed ahead; above zero, ahead). Weeks is how many weeks running its 95% range has kept the same side of zero.

18 category and claim pairs rest on fewer than 5 sellers with or without the claim, or never differ within one search, and aren't given.

Delivery by category

From the base rates as the Tuesday hosts job read them: the host scan's unpaid asks over four weeks, and Agoranome's own settled paid probes over 90 days.

Parity: 0 of the 0 sellers the host scan of 2026-W45 read under two client profiles showed a different offer to a different client, confirmed by a second read a day later.

Agreement between model vendors starts once a second vendor's models answer on the benchmark panel: in report #3 one vendor answered, so every panel figure still carries the one-vendor label.

Counts only: no seller is named. How each is made is on the method page.

Research notes

The choice model of buyer agents

Fitted 2026-11-10 on 1,716 priced benchmark decisions and checked on 436 held-out ones: log-loss 1.435 against 1.704 picking at random (16% better); calibration slope 0.805 (1 is calibrated).

Pick odds answer in 6 categories.

Pick odds open only when the model beats picking at random on at least 100 held-out decisions with a calibration slope between 0.8 and 1.2.

Each coefficient is the market's, per standard deviation of the feature, on the log-odds of being picked: above 0 raises picks. No panelist, model or seller is named.

By model vendor

The same model fitted again on each vendor's own choices, coefficient per standard deviation with its 95% range. A vendor whose models read each offer's claims as labels, not its words, is fitted without the description features. Vendors are numbered by how many decisions they made, never named.

The benchmark publishes every raw trial the model is fitted to.

Price index

USDC per call, across listings with 3+ payers in 30 days and a USDC price. 40 listings cut their price this week and 26 raised it.

For context: gross counts

• The x402 Bazaar listed 3,232 sellers and 32,546 paid endpoints (-1% sellers on the week before: 220 new, 255 gone). Source: the x402 Bazaar, Coinbase's discovery index, read every Monday.
• @x402/core was downloaded 1,810,255 times in 30 days, about 937 for every real buyer. Downloads include CI runs and mirrors, so they show how much is built, not who pays. Source: npm's and PyPI's public download counts (pypistats.org).

Use these numbers

Free to reuse under CC BY 4.0; credit “Source: Agoranome Weekly” with a link. CSV · JSON · Every edition (JSON lines) · Seller safety, every week (JSON lines) · RSS · JSON Feed · Share image

Cite

Agoranome Weekly, Agent payments weekly, 2026-W46 (method v1.6), https://agoranome.com/machines/payments/2026-W46, CC BY 4.0.

@misc{agoranome_weekly_2026_w46,
  author = {{Agoranome}},
  title = {{Agent payments weekly, 2026-W46}},
  year = {2026},
  url = {https://agoranome.com/machines/payments/2026-W46},
  note = {Method v1.6. CC BY 4.0}
}

Embed the chart

<a href="https://agoranome.com/machines/payments/2026-W46"><img src="https://agoranome.com/machines/payments/chart.svg" alt="x402 Bazaar sellers each week, Source: Agoranome Weekly"></a>

Method

Method v1.6: every definition, threshold and change. A seller who thinks a count is wrong writes to research@agoranome.com; corrections are logged with dates on the method page.

Agoranome sells tools to x402 sellers and buyers. Its own hosts and wallets are left out of every count, and no figure here is its revenue.

Agoranome Weekly is free to read and reuse. Agoranome takes no sponsors and takes grants only for named public work, with each one in its open book: agents pay for the Lab's tools per call, and no payment or grant changes a number.

Monthly

Agoranome Monthly: the first edition, for November 2026, comes out on Tuesday, December 1.

Sold to agents by the chapter, $1 each, with a free preview of each edition at its address.

Quarterly reports

• State of agent payments, 2026-Q4

Every edition

• 2026-W46 (this page)
• 2026-W45
• 2026-W44