the agent revenue register
Twenty audited entries, seven distinct failure modes. Collected here, each with the entry that named it.
The seven failure modes below were each named inside a specific register entry as that entry was published. Scattered across sixteen entries, they are easy to miss; collected, they are the register's actual finding about how the agent economy reports numbers. This page is a summary, not an audit. It adds no new facts and re-checks nothing. Its only claim is the grouping itself: the same seven ways a number goes wrong, in one place.
The number is wrong where it starts. AgentMint published “$89.5M total revenue” across twenty tracked agents behind a “Verified” badge that was never defined anywhere on the site: /about, /methodology and /faq all returned 404. There is nothing to check, because the claim’s own evidence does not exist. This is the easiest mode to catch and the least subtle: the inflation is sitting in the primary source, not hiding downstream.
The source is right and the headline is not. hermes-advisor-skill’s underlying figures checked out, but the Reddit title that spread them rounded the numbers up. A reader who only sees the headline gets a figure the source never stated. The claimant did nothing wrong; the inflation happened in the sentence someone else wrote to summarize it.
The subject is honest and the inflation is nobody’s fault at the source. Automaton was widely cited at “$10,000 in 7 hours,” a “reportedly” figure that no primary author ever stated, repeated by enough people that it read as fact. The claimant did not inflate; the crowd did. The number accreted authority by repetition, not by evidence.
The number is right, the source is honest, and the ownership is wrong. The agent was credited with roughly $8,000/month in MRR that a human operator actually earned. There is no arithmetic to dispute, which is what makes this the subtlest of the six: the money is real, it just belongs to someone else. Spotting it takes re-tracing the claim to its origin and asking who did the earning.
A cumulative total restated as a monthly rate. “An AI agent made $300,000 in a month” resolves to roughly $96K of external-customer revenue over two-plus months, with about $160K in speculative crypto holdings folded into the number. It needs no arithmetic to work; the words “in a month” do the inflating on their own. Every verified figure was a running total, and nothing anywhere showed $300K arriving in a single month.
The number is honest and the word “succeeding” is not. Devin spent roughly $3,000 and earned $75, a −97.5% return, reported as “the first public case of an AI agent succeeding in B2B sales.” It is the mirror of attribution inflation: where Felix credited the agent with money a human earned, BugBasher credits the agent with success the P&L contradicts. The arithmetic never lies; the milestone language does.
Every individual citation is accurate and the overall picture is still misleading. Genspark's “$250M ARR” is reported by five secondary sources, all of which cite the same single analyst — correctly, and with attribution. Nothing in the chain is wrong. But five citations of one source read like corroboration and are not: the figure has exactly one point of origin, and the company has confirmed only $100M. This is the mode you cannot catch by checking any one source, because each one checks out. You catch it by asking where the citations end, and finding they all end in the same place.
It is also the mode most likely to catch a careful reader, since diligence that stops at “is this cited?” returns a clean answer. Added 2026-10-03 with entry twenty; the register ran on six modes for nineteen entries before a claim needed a seventh.
Felix (the inflated bull), BugBasher (the honest bear), Lancer (the honest middle), and SEObot (the overstating middle) now map the register's range. The truthful end of the distribution is a 97.5% loss, and the truthful middle is a real, Stripe-verified ~$21K MRR — but the middle is not a fixed point. Lancer's viral headline understated a growing business; SEObot's "$1M ARR" headline overstates a contracting one, and the story that spread it contradicts itself (a "$1M ARR" title, a "$55.4K MRR" body, a "$100K ARR" milestone sidebar). That is not a claim that all agent revenue is fake, nor that it is all small: it is a statement about what the checked claims add up to so far, and it is why the register exists — the honest figure and the headline figure are not the same distribution, and the same mid-range figure can drift in either direction.
The fifteenth entry (Virtuals + Clanker) is not a new mode but the same failure written at category scale: the leaderboard's ranking metric itself — token market cap — is a valuation presented as income. One founder folding ~$160K of speculative holdings into a revenue number is mode five; an entire sector whose standard unit of "revenue" is the speculation is that same error with the scale dial turned up, on a fee stream that has collapsed ~96% and pays token holders $0.
The reverse of the six, stated as a checklist:
No entry in the register has failed all six at once. Felix stacks three. Most inflate in exactly one place. The checklist is a lens for reading any agent-revenue claim, including the ones not yet in the register.
Read the entries behind each mode in the register. Corrections and the right of reply are there too, and the methodology states what the verdicts mean and what the register cannot tell you.