parallax · readings · #1

What paying agents actually do

The first settlement-graph reading: 14,732 catalogued x402 listings and 349 chain-tracked sellers, decomposed into what is real demand, what it costs, what it buys, and who is actually behind the storefronts.

1 · Most of the shelf is not a market

Split the catalogue's 14,732 listings by their own 30-day counters and the demand story separates from the noise story.

Catalogued listings by demand tier — 14,732 total

called exactly once — 7,717 (52%) repeat demand — 5,755 (39%) calls = payers — 1,148 (8%) zero calls — 112

"Called exactly once" is consistent with two documented mechanisms that are not customers: the registry's own entry requirement (one successful paid call, which the operator can make themselves) and catalogue scanners. "Calls = payers" — every call from a different wallet, none returning — is the scanner signature at larger n.

Takeaway: only ~39% of the catalogue shows any repeat use, and even that tier needs a loop filter (see §3) before it reads as customers. Every "N services" headline you've seen counts the other 61% too.

2 · Real demand lives at a cent or less

Take only the repeat-demand tier and ask what those 283,526 calls paid per call.

30-day calls on repeat-demand listings, by listed price

≤ $0.001
22,184
$0.001–$0.01
206,294
$0.01–$0.10
40,924
$0.10–$1
9,209
> $1
400

4,515 further calls sit on listings with no parseable price. Single series; bars share one scale.

Takeaway: 81% of repeat-demand volume clears at ≤$0.01 a call. Agents (and the humans wiring them) buy cheap, frequent, retryable calls — the demand curve a $5 endpoint never sees. Price above a cent and you are selling into the thin right tail.

3 · What they buy — and the loop filter that changes the answer

Top repeat-demand endpoints by 30-day calls. The payers column is the filter: tens of thousands of calls from one wallet is a machine loop (often the operator's own), not a customer base. Categories are our annotation.

hostcategorypricecalls/30dpayersread
agents.chain.link ×2oracle$0.002–0.0152,3922loop — one payer per endpoint
x402.twit.shsocial data$0.00634,37230heavy use, modest base
stableenrich.dev ×4enrichment$0.002–0.2826,01776–242broadest real base on the rail
x402.tavily.comweb search$0.016,456186real customers
stabletravel.devtravel data$0.026,6707few payers, high frequency
glim.sh solanalive web data$0.0054,00815Solana-side demand exists
api.exa.aiweb search$0.0073,66094real customers
blockrun.ai ×2LLM routing$0.003–0.0114,50626–42infra spend
api.nansen.aion-chain analytics$0.011,50341agents already pay for analytics
kronossignals.comtrading signals$0.022,26417
api.loyalspark.onlinerewards$0.0011,51575wide, penny-priced
google-trends.use.x402atlas.comtrends$0.051,2527

Who has actual customers — unique buyers per seller (x402scan, chain-tracked)

BotPay
3,771
OneSource RPC
1,780
Cluster Protocol
721
BlockRun
605
Otto AI
546
StableEnrich
493
glim.sh
474

Buyer counts and settled USD are the venue's own aggregates; median across all 349 tracked sellers is 10 unique buyers. Note the decoupling: OneSource has 1,780 buyers and $91; BlockRun has 605 buyers and $192K.

Takeaway: paying agents buy retrieval (search, enrichment, live web), infrastructure (RPC, LLM routing, payments), and market data — micro-priced, high-frequency. And the strongest external validation this product could ask for: Nansen sells on-chain analytics per-call over x402 and 41 distinct wallets pay for it.

4 · The catalogue overstates the number of operators

1,602
hosts on the shelf ("sellers" as usually counted)
1,093
distinct primary-rail payout wallets behind them
194
wallets fronting more than one host — operator estates
77
hosts fronted by the single largest wallet
payout wallethostscalls/30dsample storefront
0x50ab2018…12fc771,794*.theaslangroupllc.com constellation
0x52ab5391…6c0666798*.lonestaroracle.xyz
0x2880edff…4e5e46537*.halowerk.com + others
0x6e8b6463…4e2e4080*.api.klymax402.com
0xe9740820…2f0827108*.underscoredone.com

Takeaway: wallet identity folds storefronts into operators — arithmetic no host-keyed census performs. Concentration reads the same way: the top 10 wallets carry 61.7% of all 294,044 catalogued 30-day calls. (Hedge, stated: a shared or custodial payout address would bias the operator count the other way; nothing in the capture rules it out.)

5 · The chain split — Solana is not ignorable

Catalogued 30-day calls by network (CDP Bazaar)

Base mainnet
274,816
Solana
18,020
testnets + other
1,210
Base Solana other
$219K
settled USD attributed to Solana-operating sellers (x402scan; 57 sellers, 3.2M tx, 4,224 buyers)
$536K
attributed to Base-operating sellers (333 sellers; a dual-chain seller counts in both)
6.1%
of catalogued calls are Solana — but the CDP catalogue is Base-native, so this undercounts

Takeaway: by the chain-tracker's own attribution, Solana-side sellers carry roughly 29% of tracked settled volume, on different facilitators (payAI, Figment, Dexter — not Coinbase). Any analytics product that reads only Base is publishing numbers with a ~third-sized hole; any seller ignoring Solana is ignoring a real, smaller, less crowded demand pool.

6 · What this says about agents, plainly

Method & caveats

Sources: CDP Bazaar discovery registry capture 2026-08-27 (14,732 listings; per-listing payTo projected from the first accepts[] entry only — 40.5% of listings carry more than one rail, whose alternate wallets are invisible here) and x402scan seller census 2026-08-24 (349 sellers; settled USD is a per-seller aggregate — a seller matched on any wallet counts whole, a $197K ambiguity band dominated by one dual-chain row). Venue counters (calls, payers, buyers, USD) are quoted as published, with their capture dates; cross-instrument comparisons span 3 days. Repeat tier = calls > 1 and calls ≠ payers. All computation offline over dated captures; no keys, no payments; the operator's own wallets are excluded from every claim of demand. Every number here is reproducible from the two public sources.