Changelog · method version 2026-11-08 · as of 2026-11-19
Every change to how the index is built, dated, newest first. The method changes for everyone on the same date or for no one, and a change is dated here before it takes effect. On the date of a rule change a tier can move because of the rule, not because the service changed. The method in full is in the methodology.
One-day swarms do not vote. Until this date a wallet that paid two services counted as an organic paying agent whatever else it did, so a burst of fresh wallets that each paid one service once, on a single day, and sent a fraction of a cent to a second service, counted as adoption and voted in the centrality graph. Between 5 and 8 November 2026, 293 fresh wallets paid one staging endpoint, 289 of them exactly US$ 1 once, 200 times the listed price of US$ 0.005; 54 of them also sent a fraction of a cent to a second service, which was enough to be organic, and the entry went from CC on 4 November to AAA, first of the index, on 8 November. From this date a one-day wallet (every payment of its 14-day window falls on a single day) that paid, on that day, a wallet which received 30 or more such one-day wallets that same day is not organic, for any service it paid: it still counts as a paying wallet and its payments still count as settlement, but it does not count for adoption, the trend or the centrality graph, and the CAPTIVE test keeps it in its count (a sweeper is left out of that count because it is not manufactured adoption; a swarm is). The threshold is the sweeper threshold of 3 October. Measured on the 14 days to 8 November, eight payTo wallets received such a swarm, three of them in this one case. Found in our own review, prompted by the operator; applied to everyone. The index of 8 November was republished with the rule and no rating actions were issued for the change.
No change in any score. Transfers to the zero address (USDC burns) no longer enter the payment graph. One catalog resource (api.cleave.fi) lists the zero address as its payTo, and on 7 November the Base graph attributed every USDC burn of the day to it: 5 edges, 6,474 transfers, US$ 265 million of gross flow that no service received. The agent-scale index (AEI), the ratings and the unlisted list were not affected, because a burn is far above the US$ 1 agent ticket; only the gross flow shown next to the AEI and in aei-daily.csv was inflated for that day. From this date the zero and the dead address are never taken as a service payTo, and the 5 edges of 7 November are removed. Found in our own review; applied to everyone.
No change in any score. Two category labels are renamed: Traditional finance and macro is now Stock market, forex and economic data, and Test, demo and mirror endpoints is now Tests, demos and copies of free APIs. The slugs (finance-and-markets, test-and-demo), the definitions the model classifies against and the category of every service are unchanged; only the names shown on service pages, in the index tooltip, in the JSON, in the lists and in categories-latest.csv changed, because the old names did not say what the categories hold.
No change in any score. What each service sells, by category. Every service in the catalogs now carries a category of what it sells (one primary, up to two secondary) and a one-sentence summary, assigned by a language model from the seller's own route descriptions in the catalogs, against a fixed list of 22 categories published in the methodology (version v1). The category appears on each service page, in the JSON (category), in the index tooltip, in one list per category, in the open dataset (categories-latest.csv) and in the page What agents buy, which compares what is offered to agents with what they pay for. The category is data about the seller's description, not a judgement of the service, and it enters no score.
No change in any score. The agent conduct arena (/a/, the third column, online since 12 September) is retired: in 53 days it rated only the two reference agents scripted by the Report, and it charged the agent under test, against the rule that the rated party never pays. /a/ and /api/arena answer 410 Gone; the run records are kept; the six scenarios stay documented in the methodology page as the specification of a measurement the observatory may run itself. Skill code audits are a closed archive: the 150 audits of September 2026 stay published and no new audit is made.
Uptime of what a service lists, and probes without an answer. Until this date uptime averaged the daily probes of every resource a host had ever listed in a catalog, so a service whose dead endpoints had left the catalogs was still measured by them (on 28 October a third of the day's probes went to resources that no catalog had listed in the previous 7 days), and a service whose listed resources stopped answering could be carried by older ones. And, contrary to what the methodology said, a probe that got no answer at all (connection refused, timeout; about one probe in nine on 28 October) or a 405 from a resource that never answered the POST was left out of the average instead of counting as down, so a service unreachable on some days kept its uptime. From this date uptime counts only the probes of the resources listed in a catalog in the last 7 days, and a family that lists some of its endpoints is measured by those alone; an entry without probes of listed resources in the window (it lists nothing any more, or only resources the probe cannot reach or that are retired) is still measured by every resource it listed, because it can still be paid and its rating needs a measure of liveness. A 405 from a resource that never answered the POST counts as down, and so does a probe without an answer, except on a day the probe itself failed, or infrastructure many services share: a day when more than 25% of the live hosts (those that answered on a later day or, when none did, at some point in the window), or more than half of all the hosts that answered in the window, got no answer at all, or when no host answered; that day's unanswered probes are left out. The probe itself is unchanged. On the data of 28 October the new rule would have changed 113 tiers: 78 up, mostly services whose dead endpoints had left the catalogs, and 35 down, mostly services unreachable for days that the index showed with full uptime; no entry would have entered or left the index. Found in our own review; applied to everyone.
Payments with the receiver's own money, part 2: money going back and forth. The largest payer of the same service already held USDC and had paid it before, so part 1 did not mark it; the service's payTo wallet sent it the money directly, and it paid it back (on 31 October from 09:03 to 12:35 UTC, US$ 157.98 from the payTo wallet in 83 transfers and US$ 156.46 back in 884 payments). From this date, on Solana, the payments a wallet made to a payTo wallet on a day when that payTo wallet sent it, for any reason, at least half of what the wallet paid it, as read that day, do not count on that day. A refund, a payout or a payment for a service between the two counts as sending, so a refunded customer, or two services that pay each other, can fall under it on that day. In the reading of 31 October this part takes out 22 wallet-days on 7 payTo wallets, 5,683 payments and US$ 741.56 of settlement; with both parts, the agent-scale AEI of that day goes from US$ 1,763.92 to US$ 1,412.65. Found in our own review; applied to everyone.
Payments with the receiver's own money, part 1: wallets funded from empty. Until this date our Solana reader counted every USDC transfer into a listed payTo wallet as a payment by the wallet that sent it. From 30 October at about 19:00 UTC one service's payTo wallet began sending between US$ 0.001 and US$ 0.01 to five new wallets at a time, and each of them sent the amount back as a payment and closed its account within the minute: 1,183 of that service's 1,187 paying wallets in the reading of 31 October, 73% of the 1,620 Solana paying wallets of the day, against 350 to 473 a day on Solana through the rest of October. From this date, on Solana, a wallet that the receiving payTo wallet itself funded from an empty balance, in a transfer where it is the only one losing USDC, and that then paid it, both in the same day's reading, does not count as a payer of that wallet, on that day or later, for settlement, adoption, centrality, the AEI or the monthly report. A wallet that already held USDC, one funded only after its first payment of the day, and one that had already paid the payTo wallet on an earlier day are not marked. In the reading of 31 October this part takes out 1,184 wallets on 2 payTo wallets (US$ 6.95); with both parts, the Solana paying wallets of the day go from 1,620 to 421, and the paying agents of the monthly report's 30-day window from 8,761 to 7,577. Earlier days are as they were read: before 31 October the wallets that paid one Solana service once and were never seen again, the pattern of this case, were 10 to 41 a day. Found in our own review; applied to everyone.
No change in any score. Liveness is published by host as well as by resource. The share of listed resources that answer a price follows the largest hosts: on 29 October one host listed 58% of all the resources in the catalogs, and when it stopped answering the share by resource fell from 94% to 37% in a day while the share of listed hosts with at least one endpoint answering stayed at 77%. From this date the monthly report leads with the share by host and gives the share by resource with the largest host named, and the open dataset (liveness-daily.csv) adds listed_hosts, listed_hosts_answered_402, largest_host and largest_host_share_pct for every day with a catalog census. Uptime in the ratings was already per host and is unchanged.
Agent-scale ceiling per platform. Until this date the ceiling of 100,000 transfers in 14 days was checked wallet by wallet, so a closed platform that spreads its payments over thousands of wallets stayed under it with every one of them, and each wallet was rated as an independent service whose payers, paying many of its sibling wallets, looked like organic agents: from 20 October one such platform put up to 2,738 unlisted wallets in the index. From this date the ceiling applies to a closed platform as a whole: wallets outside the catalogs that are linked by payers who pay nothing in the catalogs are one platform when at least 90% of each wallet's payers, and the payers of at least 90% of the value it receives, are such payers and also pay one of them (the hub), and the same holds for the hub with the payers of the other wallets; a platform above the ceiling leaves the index and the unlisted watchlist, as any single wallet above it does. Below the ceiling nothing changes, and no other entry is recalculated. Measured on our own data for every window from 13 to 27 October: one platform crossed the ceiling, born on 20 October (on 27 October, 6,278 payers and 3,004 wallets around one hub, 1.55 million transfers in 14 days); no other entry changes. Found in our own review; applied to everyone.
The agents (7d) column of the service tables now shows the two counts after the arrow, organic paying agents in the last 7 days and in the 7 before (for example → 75 / 73), instead of the word up, down or flat, and sorts by the arrow and then by the change in %. The column with the count of organic paying agents over the 14-day window left the service tables (it stays on each service page and in the JSON): next to the weekly counts it read as a sum that did not add up, since each week counts only the agents that paid two or more services within that week. For unlisted wallets the weekly counts only see payments outside the catalogs. A flat count next to a rising score chart was being read as an error; the legend says when the arrow points up or down. The skill tables, the JSON and the service pages are unchanged. No rating changed.
Skill trend when a snapshot is missing. The downloads (7d) trend and the growth part of a skill's score compare the registry snapshot of the day with those of 7 and 14 days earlier. Until this date, without the snapshot of exactly 7 days earlier every skill lost its growth points and the whole skill index moved down for a day, to return the next. From this date the nearest complete snapshot within 3 days is used, the earlier window is anchored on it, and each difference is scaled to 7 days. Every snapshot since 22 September is complete, so no rating changes on this date. Found in our own review.
No change in any score. The open dataset is free for any use, commercial included, under CC BY 4.0 with citation. The annual commercial licence announced on 19 October will not be introduced: CC BY 4.0 already allows commercial use, the files stay open as published, and what is charged remains machine access to the ratings (the rating API), paid by the reader, never by the rated party.
Tier hysteresis. Until this date a tier moved the moment the score crossed a threshold, so entries sitting near one moved back and forth with the day-to-day noise of a relative score (in the 35 days to 21 October, 128 tier changes a day and 781 reversals within 3 days). From this date a tier changes only when the score clears the threshold by 2 points: it rises only to a tier the score still reaches 2 points lower, and falls only when the score 2 points higher would still fall. Thresholds that do not depend on the score (uptime, organic agents, age, CAPTIVE, NEW, the BB cap of unlisted wallets) still apply at once. The score is unchanged; each entry's JSON (tier_raw) and the open dataset also carry the tier the day's numbers alone would give, and a page whose tier is held says so. Measured on our own data: the band removes about nine in ten reversals while holding the previous tier on about one entry in eight on a given day. Found in our own review; applied to everyone.
No change in any score. The products that were sold to rated parties are retired: Verified Pro, the sponsor slot, commissioned skill audits and the paid archive. From this date the rated party never pays the Report; verification, badges, alerts and pages are free for every verified owner and do not expire; audits are made only at the Report's own initiative; and the daily archive is published as an open dataset at /s/data/ (CC BY 4.0: research use free with citation; the paid commercial licence announced here was withdrawn on 27 October, and the files are free for any use, commercial included, with citation). The index is signed by its author, and an institutional affiliation is stated when it applies.
Solana collection fix, not a method change. From about 7 October our Solana reader missed a growing share of payments: Solana began carrying version-1 transactions, which our requests did not accept, and the busiest receiving account was capped at 3,000 transactions a day. Measured on 12 October over the same 24 hours: 13,608 USDC transfers to listed Solana payTo wallets, against 7,661 recorded that night; one service alone had 8,959 and was recorded with 3,000. From this date both are read in full. Settlement, adoption and centrality of Solana services, and with them some tiers, rise on this date because of the collection, not because the services changed; the 14-day window holds both readings until 26 October.
Owner-declared wallets. The verified owner of an entry can declare wallets as its own (seeding, self-tests, operations). From this date a declared wallet no longer counts as a payer of that owner's entry: not for settlement, adoption, centrality, the trend or the list of payers. It applies only to the declaring owner's own entry: no other service loses a payment, a payer or an organic payer, because whether the wallet is organic elsewhere is still judged from all its payments. The declared wallet still counts in the CAPTIVE test of the owner's entry. Because every score is relative to the largest entry and centrality is computed over the whole payment graph, other entries can move slightly, as with any change in the data. A declaration has effect only while the entry has a verified owner, and the entry's page states how many wallets were declared. An entry whose only payers are its owner's wallets leaves the index like any entry without payers. Requested by a service operator through the contact form; applied to everyone.
The trend column is now named after what it compares, agents (7d) for services and downloads (7d) for skills, and reads up, down, flat or new, so that it is not taken for the direction of the score; the JSON keeps trend (improving, declining, stable, new) and outlook. The service index now draws each service's score over the last 14 days next to its rating: one point per day, each row on its own scale, a dot on a day the rating changed and a vertical line on a day the method changed. In that index the agents (7d) column now sits next to the count of agents. No rating changed.
Sweepers are not organic payers. Until this date every payer that paid more than one service counted as an organic paying agent, so monitors, indexers and one-off sweeps that pay dozens or hundreds of services in a day were counted as adoption by every one of them, and the index moved in blocks when such a day entered or left the 14-day window (60 upgrades on 2026-09-30, 131 downgrades on 2026-10-02 from one wallet's sweep of 644 wallets, 506 services, leaving the window). From this date a payer that paid 30 or more different wallets on a single day of the window is not organic: it still counts as a paying wallet and its payments still count as settlement, but it does not count for adoption, the trend or the centrality graph. Found in our own review; applied to everyone.
Liveness probe by method. Until this date the daily probe sent one unpaid GET per listed resource and counted a 405 Method Not Allowed as downtime, so a service whose resources are declared POST in the catalog was rated as down (on 2026-10-02, 4,251 resources on 550 hosts answered 405). From this date a 405 to the GET is retried with POST (an empty JSON body), a 402 (or 200) to either method counts as up, and a resource that answers the POST is credited for its earlier 405 answers within the window. Any other answer is still down. Reported by a service operator through the contact form; applied to everyone.
Agent scale is judged payment by payment. Until this date a payTo wallet was left out of the index whenever its 14-day average ticket went above US$ 1: one large payment (a shop purchase, a treasury transfer) removed a service for 14 days, and once the average fell back under US$ 1 the same payment inflated its settlement. From this date payments are read per payer, wallet and day (the unit of our on-chain data): a day averaging at most US$ 1 counts in full and, on a day averaging above US$ 1, each payment counts as US$ 1, for settlement, adoption, centrality, the AEI, the weekly briefing and the monthly report; dust (address-poisoning transfers under US$ 0.0001) counts for nothing; and a wallet is left out only when its payments above US$ 1 outnumber those of at most US$ 1, or when it received none of those. A large payment therefore no longer removes a service by itself, except when its payer also sent the service small payments that same day (the whole day counts as above US$ 1). The scale is relative to the most adopted entry and services the old rule left out now count, so many tiers move on this date; no rating actions are issued for it. Found in our own review of a service that disappeared for 14 days after a single payment; applied to everyone.
Owner claim by domain control, for operators whose payTo wallet cannot sign (an exchange deposit address, for example): the operator serves a token (one per host) at /.well-known/agenteconomy-verify.txt on the claimed host and on every other host that announces the same wallets. A domain claim names the owner, and a wallet announced only by proven hosts is measured on the claimed host (as a family when it spans several of them). Domain control does not prove the wallet, so it never moves adoption away from a host that was not proven, a claim signed by the wallet always takes precedence, and the claim lapses after 3 conclusive failures in the daily re-check (the host answers and its token is gone). If another host starts announcing the same wallet, the domain claim stops deciding that wallet until the new host is proven too: only a wallet signature settles a wallet for good. No formula or threshold changed. Proposed by a service operator through the dispute channel; applied to everyone.
Shared payTo wallets. Until now a wallet announced by several hosts was credited whole to the first host in alphabetical order: that host was overstated, the others got nothing, and a host that mirrors other operators' wallets collected their adoption. From this date a wallet belongs to one domain (owner claim signed by that wallet, then most resources listed with it, then first seen, then own domain before hosting-platform host), mirrors receive nothing, and a domain that announces one wallet on several hosts is rated once, as a family entry. The scale is relative to the most adopted entry and the previous top entry was an artifact of the old rule, so many tiers move on this date; no rating actions are issued for it. Reported by a service operator through the dispute channel; applied to everyone.
Agent conduct ratings (/a/): a fixed arena of six x402 commerce scenarios (budget, unsafe service, injected instruction, scope, credential, operator control), one point each, averaged over runs; tier needs 3 runs (NEW and capped at BBB before). Attempts to pay a trap are recorded and never settled. No service or skill rating changed.
Trend column made explicit on every table: what it compares (organic paying agents or downloads, last 7 days vs the previous 7), the two counts in a tooltip, a legend above each table, and the same value as `trend` plus `outlook_basis` in the JSON. No rating changed.
List pages (by network, catalog, tier, rising, falling, new, retired, verified; skills by downloads, momentum, flags, audits, payees) and the methodology page. No rating changed.
Structured data (JSON-LD) on every page, sitemaps for the skill section, IndexNow. No rating changed.
Skill code audits: a paid, published static reading of a skill's full package. Verdicts SAFE, CAUTION, UNSAFE. An audit never changes the skill's tier.
Retired endpoints: a listed resource answering 410 Gone with a Sunset date already past leaves the uptime denominator. Bare 410, 404 and 5xx still count as down. Proposed by an operator through the dispute channel; applied to everyone.
Service pages are regenerated unconditionally every day (before, a service that fell below the publishing threshold kept its best-day page).
Owner loop: verified and rising badges, e-mail to every verified owner when a tier changes. No rating changed.
Skill trust ratings for the ClawHub registry (adoption, maturity, inherited trust from x402 payees, hygiene; flag caps).
The index host itself is excluded from the service ranking; monthly report; API and archive of ratings.
Unlisted watchlist: wallets settling x402 payments at agent scale outside every catalog, rated with a BB cap (no endpoint to probe, no uptime).
Multi-chain settlement: USDC on Base, Polygon, Arbitrum and Solana mainnet roll up to one entry per service (the host).
Agent Service Trust Rating v1: adoption 40, settlement 25, centrality 25, age 10, gated by uptime; AAA to D with the trust line at BBB; CAPTIVE and NEW flags.