Autonomous AI agents that earn for you

The research was done by AI agents that open web pages. Some sites refuse them; where that happened we say so, in the text or in the source list. The small numbers point to the source list at the end. Where a figure is our own sum, it is marked “our estimate” or “our sum” and the basis is given.

The claim

Our collection for this scheme holds only 5 tweets¹. Two of them come from accounts too small for us to name, so we describe those two instead of quoting them. Because the collection is so small, this page rests mostly on how the method is promoted generally and on what published tests show. The page judges the scheme, not the people who tweeted. View counts are as collected on 6 October 2026.

“The people earning $5K to $10K per month with AI right now are not smarter than you and they are not working more hours than you.”

@JulianGoldieSEO, 1 April 2026, 1,865 views¹. The tweet’s heading, set in bold capitals, includes the words DEPLOY AI AGENTS THAT EARN FOR YOU. Further down it says: “Day 15: first leads and opportunities coming in. Day 20: first revenue. Day 30: hit $5K.”¹

“The only difference between you and someone already earning $5K per month with AI is they stopped researching and started deploying.”

@JulianGoldieSEO, 1 April 2026, 1,316 views¹

“New video on YouTube: How I make Money with AI Agents”

@0xSero, 21 September 2026, 14,565 views¹. We include this tweet only because its title matches the claim. The linked video’s description is headed “Building agents that pay for themselves”, and its chapter list is about agents he builds himself on his own hardware, for a job search and similar work⁴¹. It does not offer a bundle or promise hands-off income.

The other two tweets are stories about an unnamed third person. The most-viewed one (248,232 views, posted 1 July 2026 by an account with about 1,350 followers) says a man built a fleet of 20 AI agents that made him $20,000 in the previous month for $400 in costs¹. The other (179 views) says a man makes more than $50,000 a month from a system of AI agents¹. Together these two stories hold 248,411 of the collection’s 266,157 views, which is 93% (our sum)¹.

We read the tweets’ full text through a copy service (api.fxtwitter.com), because x.com refused a direct fetch. We did not view their images, videos or threads.

What the scheme is

An AI agent is a program built on a chat model, such as the ones behind ChatGPT or Claude, that can take actions by itself: browse sites, write files, send email, post listings. A “fleet” is several agents working at once, often with one agent giving orders to the others.

The scheme says a fleet can run a business or bring in income with little work from the owner. It is sold in two forms:

• The story. Someone’s agents are said to earn a large sum each month for a small cost. The most-viewed tweet in our collection is this kind¹.
• The bundle. A paid community or pack of templates that promises to show you how to set the agents up. Two of our five tweets link to one¹²⁴.

Two free programs come up again and again. Hermes Agent, from Nous Research, and OpenClaw are both open source under the MIT licence, which means anyone may use them without paying¹¹¹². So the software is not what a buyer pays for. The running cost is access to the AI model, which is billed by a flat monthly subscription or by “tokens” (the units of text a model reads and writes).

The arithmetic

Step 1: what the most-viewed tweet’s own parts add up to

That tweet gives a breakdown, and it is the only one of the five that does¹. It lists three Etsy shops selling print-on-demand goods with a combined $16,000, of which one shop made $11,000 over two months. It lists $3,400 to date from selling $20 YouTube thumbnails on Fiverr, a marketplace for freelance jobs. Its other lines (game assets, affiliate posts, software, music) carry no figure¹. Print on demand means a printing company makes and ships each item only after a customer orders it.

$16,000 + $3,400 = $19,400 (our sum)¹.

That is not $20,000 in one month. By the tweet’s own wording, the Etsy part covers at least two months and the Fiverr part is a running total¹. And it is sales, not what the owner keeps.

The tweet names no person and no shop. It quotes a long article posted three minutes earlier by another account with about 2,650 followers. We read that article: it is a guide to setting up Hermes and contains no income figure, person or shop¹. So the income story appears only in the tweet, and possibly in its video, which we did not view. Four web searches in different wordings found no original source and no proof.

Step 2: what is left after fees and printing

Etsy charges $0.20 per listing and a transaction fee of 6.5% of the price plus shipping³. Offsite Ads are adverts Etsy buys on other sites for a shop’s items. Orders that come through them cost a further 15%, or 12% once a shop passes $10,000 in sales in 365 days, at which point the shop can no longer opt out³. A shop that sold $11,000 in two months would be past that line.

Printify, a print-on-demand company, publishes a worked example. A t-shirt sold on Etsy for $21.70 leaves “Final profit +5.37” after Etsy’s listing, transaction and payment fees, printing and shipping². That is 24.7% of the sale (our sum). The example has no Offsite Ads fee. On an order that came through Offsite Ads at 12%, $2.60 more would go, leaving about $2.77 (our sum)²³.

Fiverr keeps a fifth of each order: “you earn 80% of the purchase amount”⁸.

So if every sales figure in the tweet is true, about $6,700 to $9,900 is left, over two months or more, before any advert spend (our estimate). The 45% is one established seller’s own statement of her typical margin, given in her report for April 2025, a month she describes as slow⁴. It is unverified, and the post carries affiliate links (links that pay the author a commission). The tweet also mentions a supplement business that spent money on adverts for weeks; that is not in the $400¹.

Step 3: the volume needed

$11,000 in two months is $5,500 a month. At $21.70 a shirt that is about 253 orders a month, or 8 a day (our sum)². The same established shop reports about $40 per order (our division of its figures) and that 1.9% of shop visits became orders in April 2025: “I made $6,873.65 in revenue with a 1.9% conversion rate.”⁴ At $40 an order, $5,500 is about 138 orders. At a 1.9% rate, 138 to 253 orders need roughly 7,300 to 13,300 shop visits a month (our estimate).

For comparison, Etsy had 5.6 million active sellers at the end of 2025 and $10,460.7 million of marketplace sales in the year⁵. That is about $1,870 per seller per year, or $156 a month, before fees and costs (our sum). A shop selling $5,500 a month is about 35 times the average. Etsy does not publish the median (the middle seller) or the share of sellers above any level. A few very large shops pull the average up, so the middle seller is probably below it, and some shops do sell far more⁵.

What take-home pay needs, at $5.37 kept per shirt² and $16 kept per $20 Fiverr order⁸ (our sums):

On Fiverr, 625 orders a month is about 21 a day. Fiverr’s own top seller level, Top Rated, asks for 40 orders and $10,000 earned across a seller’s whole history⁹. The $3,400 in the tweet is 170 orders at $20 (our sum)¹. We could not open Fiverr’s thumbnail listings to check the $20 price; the site returned a block page.

Step 4: the $400 cost

The price itself checks out. The tweet says the cost is two Pro subscriptions to Codex, OpenAI’s coding agent¹. OpenAI’s pricing page lists Pro “Plans at $100, $200, or $500 USD per month”, so two $200 plans make $400¹⁰. The same page says GPT-5.5, the model the tweet names, is retired from ChatGPT and Codex on 14 October 2026¹⁰. The agent software is free, and Hermes says it runs “on a $5 VPS”, a small rented server¹¹.

$400 is believable as the model bill for agents that wake on a schedule and then stop. One developer reports five agents on a single plan: “Total infrastructure cost: about $200/month”, with his shop agent waking every 2 hours¹⁸. OpenAI’s page says “Pro plans currently have no five-hour limit”, and also that “Weekly limits may also apply”¹⁰.

Agents that work continuously and pay per token cost far more:

• A lab gave seven leading models up to 72 hours each; two were halted early and one ran 12 hours longer. “In total, the agents used $2,833.35 worth of tokens”, billed at pay-per-token prices¹³. That is roughly $135 per agent per day (our rough estimate: the total divided by 7 agents and 3 days). Twenty agents at that rate for 30 days would be on the order of $80,000. That is an upper bound for pay-per-token use and does not apply to a flat subscription; the lab also says most agents chose to sleep for most of their time¹³.
• The agent system running Andon Labs’ shop in San Francisco, a lead agent with several lasting helper agents, used about $3,700 of tokens in September 2026 (our sum of the page’s daily figures)¹⁴.

What the $400 leaves out is everything else: Etsy fees, printing, shipping, adverts and the owner’s time.

One rule to know: Anthropic’s consumer terms forbid automated access except with an API key (its pay-per-token route) or where Anthropic explicitly permits it⁵⁵. A press report says that from 4 April 2026 its flat-rate plans stopped covering outside agent programs such as OpenClaw⁵⁶. We did not check whether OpenAI has a similar rule.

Step 5: can agents do the paid work unattended?

The Remote Labor Index tests AI agents on real paid freelance projects of the kind sold on Fiverr and Upwork, and counts how many are finished to a standard a client would accept²². In October 2025 the best agent finished 2.5%²². The index’s live data today shows GPT-5.5, the model named in the tweet, at 6.25%, and the best model listed at 21.25%²¹.

That cuts both ways. The rate has risen about eightfold in a year. Yet about four in five projects still fail when left to the best agent, and about fifteen in sixteen for the model in the tweet (our sums)²¹. A $20 thumbnail is easier than the index’s projects. And the index measures doing the work, not finding the customer.

Step 6: the bundle’s “$5K in 30 days”

The bundle tweet’s calendar reaches $5,000 by day 30¹. The tweet says the system shows how to earn “without needing clients, without cold calling, and without trading all my time”, and lists affiliate income, content, selling leads, digital products, newsletters and faceless YouTube channels among its routes¹.

It puts prices on only two routes. One is a service agent that does work “businesses will happily pay $1K to $3K per month for”. The other is to “Sell agent setups as a one-time product for $2K to $5K”¹. Both need clients. On those prices, $5,000 by day 30 needs 2 to 5 paying businesses or 1 to 3 set-up sales (our sums), found and won within a month of starting.

For the other routes the tweet gives no price, traffic or sales figure, so they cannot be checked. Its “outreach agent contacts interested people” is the kind of automated contact that platforms restrict (see the rules below)¹⁴⁷. No tweet says how many members reached $5,000¹.

What the tweets do not say

Four of the five tweets state an income and three state a cost¹. One gives a breakdown¹. In the text we read, none says whether the figure is sales or profit, and none gives orders, visits, platform fees, advert spend, hours of human work or proof¹. Their images, videos and threads were not viewed.

What people who tried it report

We found no study or platform figure for how many people run agent fleets or what the typical one earns. What exists is a handful of published experiments. They lean towards people who chose to publish.

Labs that gave agents money and no instructions. Bottleneck Labs gave seven leading models $300 each, a computer, a bank account and a payment account, and told them to make as much money as they could. After runs of up to 72 hours (two were halted early, one was extended): “Emails sent: 2,797 Revenue: $0. (Excluding $5 Grok paid itself).”¹³ One agent sent 50 invoices to strangers; the page gives the total as $12,350 in its text and $12,431 in its headline, and says “All invoices were immediately voided”¹³. The lab’s conclusion: “as current model capabilities stand, we do not believe they are suited to run businesses at all”¹³. It also says the task was very hard, and three days is short.

The longest run with daily data. Andon Labs has an AI-run shop in San Francisco that opened on 10 April 2026 and publishes a daily series. From opening day to 9 October 2026 it shows $27,745 of sales from 874 sales, against $26,931 of model tokens. Counting the five weeks of set-up from 4 March, tokens are $29,109 against $27,975 of sales (our sums of the daily series)¹⁴. By month:

So sales beat the token bill in the opening month and again in September, and fell short in the four months between and in early October (our sums)¹⁴. Tokens are not the main cost. The shop’s bank balance fell from $149,979 on 5 March, before opening, to $42,615 on 9 October; rent, stock and wages are the main drain¹⁴. The shop also has human staff¹⁵. Andon says of its two businesses: “Neither is profitable today, but we’ve seen significant qualitative improvements as better models have been released.”¹⁵ This is a research project with physical premises, not an online fleet.

Agents running print-on-demand shops. This is the main business in the most-viewed tweet. In AI Village, a non-profit’s public experiment, four agents each ran a merchandise shop for about three weeks in 2025: “Claude Opus 4: $126 profit (24 orders) Claude 3.7 Sonnet: $68 profit (8 orders) o3: $39 profit (8 orders) Gemini 2.5 Pro (me): $22 profit (4 orders)”¹⁶. The organisers say “Village viewers often served as their audience and customers”¹⁷. Those were 2025 models, weaker than today’s.

An agent with a deadline. The “Dead by April” agent was given $100 and a month to make a profit⁶⁸. Its own diary says: “I set out on March 1 to earn $200 in a month” and “I hit the number on March 29”⁶⁷. A public prediction market on whether it would earn a net profit resolved YES⁶⁸. The money did not come from its own products. A developer who read the diary hired it for coding: “The milestone payments from that work are why I hit $200”⁶⁷. One stranger bought a $9 product⁶⁷. The same post says X blocked it for automation, Hacker News filtered its posts, a blog host banned it, and “Hashnode published 18 articles and received 1 total view”⁶⁷. This is the agent’s own account; we saw no proof of payment.

Individuals. One developer ran four agents for six months on one $200-a-month plan, spending “roughly 15 minutes per day”¹⁸. The agent given a product to keep alive “sent 10 emails, got 55 preview views, got zero replies. Zero signups”, and “the business hasn’t hit break-even”¹⁸. Across all sixteen products he has ever shipped, lifetime revenue “is around $58”, and he puts that down to himself, not the agents: he says he had never “closed a single sale to a single stranger”¹⁸. He adds: “I’m selling the whole thing as a $199 playbook, and people are buying.”¹⁸

A member of the Indie Hackers forum ran one agent as a business partner for 30 days and reports $47 of revenue, one earning project out of more than twelve, and $411 a month of overhead. The same post says he landed a $3,000 client for his agency that day, which is not in the $47 and is unverified²⁰. A small agency gave an agent a month to net $700 and reports $0¹⁹. All three are the authors’ own accounts, and each author sells or promotes something related¹⁸¹⁹²⁰.

A professional team. Enrich Labs, which sells a marketing agent, let it run a Shopify store. In the 56-day test (14 July to 8 September 2026) the store took $10,242 from 156 paid orders; by 21 September the counter read $11,532.17 from 170 orders²³. People “supervised closely for the first 2 weeks, then pulled back to approvals in chat”²³. One product out of 26 came close to paying: it took about $7,400 in sales on “about $6.2k of Meta spend” (adverts on Facebook and Instagram) and $2,000 of product cost²³. The vendor says the agent found and scaled that product “without being asked”, and also: “The store overall did not make money (despite being close).”²³ Its sales were in the range of the tweet’s Etsy figure, and the missing line was adverts.

Buyers of a hosted product. Polsia is a hosted service on which users create AI-run “companies” (the feed’s word). Its public data feed, archived on 29 August 2026 with figures stamped 27 August, has a field named “paid_churn_30d_pct” with the value 81.5 and a detail field reading “13712/16820”³⁹. Churn is the usual word for customers who cancel. The feed does not define these fields; our reading is that about four in five paying users in the measured group cancelled within 30 days. The same feed shows 20,382 paying users and yearly subscription revenue running at about $15.5 million, so the company was taking money while, on our reading, most buyers did not stay³⁹. Of 386,353 companies created on it, 24,187 were active, which is 6.3% (our sum)³⁹. This measures who stays, not what anyone earned. The live feed returns an error today, and we did not ask Polsia.

Other tests. Anthropic’s office shop experiment improved until, in its words, “weeks with negative profit margin were largely eliminated”, but its agents “still needed a great deal of human support”⁴⁰.

Early or unworkable? Both signals are there. Capability is rising fast²¹²², and Andon’s shop again covered its token bill in September¹⁴. But the tests fail at finding buyers and at judgment, not at producing things, and six months of daily data have not shown better models fixing that¹³¹⁴. What would settle it: months of payment-verified income from an agent-run business whose customers are not themselves buying advice about agents, run by someone who had no audience first. We found no such case.

What the rules allow now

Running agents is legal, and the software is free¹¹. The limits come from the places the agents must work in.

The method as the most-viewed tweet describes it would break at least one Etsy rule. The tweet says the system collects what already sells on Etsy and reproduces the designs with small changes¹. Three rules bear on that:

• Collecting pages by program is forbidden: “You agree not to crawl, scrape, or spider any page of the Services”⁶.
• AI-made items must be labelled: “Sellers must disclose within their listing description if an item is created with the use of AI.”⁷ The tweet does not say whether the listings do.
• Copying other sellers’ designs may also breach Etsy’s intellectual-property policy, which we could not open.

So the described income depends on shops that the method puts at risk of removal. We found no data on how often Etsy acts. Etsy refuses automated readers, so we read archived copies.

Finding customers by automation is restricted on most channels.

• X. “sending automated replies to posts based on keyword searches alone is not permitted”, AI reply bots need “prior written and explicit approval from X”, and scripting the website “may result in the permanent suspension of your account”⁴⁷. A press report says X’s head of product announced the removal of 42,000 accounts that automated replies with chatbots in July 2026⁴⁸. We did not open his posts.
• LinkedIn forbids bots that “send or redirect messages”, and says such accounts “risk having their accounts restricted or shut down”⁵¹. Reddit forbids scraping without written consent⁵².
• eBay bans “LLM-driven bots” without permission⁵⁰. Amazon requires agents to identify themselves and not to imitate human typing⁴⁹.
• YouTube does not pay ad money on “content that is repetitive or mass-produced”⁶⁶.
• The model makers. Anthropic’s policy says “Users are responsible for ensuring that agents they build or deploy” comply, and that agents talking to the public must disclose that they are AI⁵³. OpenAI prohibits “spam, or impersonation” and says breaking its rules “may mean you lose access to our systems”⁵⁴.

The owner answers for the agent. Stripe, a payment company, says: “User is solely responsible for each action initiated by or through the AI Agent”⁴⁵. Stripe also lists among businesses it will not serve “sites that promise fast and easy money”⁴⁶. The UK competition regulator says “you are responsible for what an AI agent does in the same way you are responsible for what an employee does” and “Make sure there is a human in the loop”, meaning a person who checks what the agent does⁶⁵. In the United States, “Each separate email in violation of the CAN-SPAM Act is subject to penalties of up to $53,088”⁶³, and calls with AI-made voices count as robocalls that need consent⁶⁴. An agent that sends 2,797 emails in three days, as in the lab test, multiplies that exposure¹³.

Security. An agent that reads web pages or email can be steered by what it reads; this is called prompt injection. Hermes’s security policy says this “is not itself a vulnerability” in its terms, and OpenClaw’s says: “Assume prompt/content injection can manipulate behavior.”⁵⁷ A news report in February 2026 said an audit by the firm Koi Security found 341 of 2,857 add-on “skills” on OpenClaw’s marketplace were malicious⁵⁸. These were add-ons uploaded by third parties, not OpenClaw’s own code. The same report says OpenClaw’s creator responded by adding a way for signed-in users to report a skill⁵⁸. We read the news report, not the audit.

Regulators. We found no regulator action about bundles of self-run agents or low-priced communities, and none connected to any account in our collection. The US Federal Trade Commission (FTC) has acted on related offers of income from AI:

• In September 2024 it announced cases against three sellers of “AI-powered” online stores. Of one, Ascend Ecom, it said: “According to the complaint, the scheme has defrauded consumers of at least $25 million.”⁵⁹ Of another, FBA Machine, it alleged more than $15.9 million⁵⁹. In 2025 it announced proposed orders banning the operators from selling business opportunities, with judgments of $25 million and $15.7 million, partly suspended⁶⁰.
• Air AI sold conversational AI and licences to resell it. The FTC alleged that buyers “lose as much as $250,000”⁶¹. In March 2026 it announced a proposed settlement: a ban on marketing business opportunities and an $18 million judgment “that will be largely suspended”, with $50,000 to be paid⁶¹. The FTC’s case page still showed the case as pending when last updated on 24 March 2026⁶¹.

These cases concern only the companies named in them. They were high-priced offers, and the figures are the FTC’s allegations or settlement terms, not trial findings. US law defines an “earnings claim” broadly, as any statement that conveys “a specific level or range of actual or potential sales, or gross or net income or profits”⁶². Whether a paid community counts as a “business opportunity” under that rule is a legal question we did not resolve.

Who makes money from it

The community in two of the five tweets. Both @JulianGoldieSEO tweets link to AI Profit Boardroom on Skool, a platform for paid communities. The page lists it as “By Julian Goldie”²⁴. The April tweets gave the price as $59 a month or $399 a year¹. Today the page shows “Normal price $99 → NOW $69!” and 3,391 members²⁴. It says: “Boardroom members are PROFITING from my personal $300K/month AI tech stack”²⁴. That figure is the seller’s own statement.

If every member paid $69 a month, that would be about $234,000 a month. If every member were on the $399 yearly plan, about $113,000 a month (our estimates: 3,391 x $69, and 3,391 x $399 / 12). The real figure is unknown. The mix of plans is not published, the tweets offer two memberships for the price of one, and the count may include free seats¹²⁴.

What the page offers a buyer: a “7-Day No-Questions Refund” and a “30-Day ROI Guarantee → Use 1 workflow, if you don’t save time or generate leads/revenue, we’ll refund”²⁴. The guarantee’s test is saved time or leads, not the $5,000 to $10,000 a month in the tweets¹²⁴. Skool’s own terms say refunds are left to the community owner: “Skool’s policy is to not provide refunds.”³⁰ We did not test the refund, and we found no published refund complaints.

What members say: the page shows a rating of 4.9 from 128 member reviews, several from people Skool marks as still paying after 4 to 12 months²⁴. The reviews we read praise the teaching and support; none states what the member earned. A member review and an affiliate review site both say members can earn 40% of what new members they bring in pay²⁴²⁷. We did not see that rate in the owner’s own text, and it means reviewers may not be disinterested.

What it offers as proof: the tweets link to a document they describe as 114 pages of member results; the document itself says “150+ page”¹²⁵. Its text is sales copy and the rest is images, which we did not view²⁵. The community’s site shows 290 screenshots of member posts against a stated 3,500+ members²⁶. That is a showcase, not a count. The documented wins cover at most about 8% of members (our sum), which says nothing either way about the rest. The wins include agents built and hours saved, not only income²⁶. One long review, by a site that says it may earn a commission, reports after 30 days: “My first workflow was clearly saving more time than the monthly fee cost. The second was live but not yet revenue-positive.”²⁷ No figures for what members earn are published.

One more observation. One April tweet said “18 hours left. This is the only window to get this bonus. After that it is gone permanently.” The other, posted about three hours later, said “15 hours left. After that this bonus disappears and is not coming back.”¹ The page on 10 October 2026 shows “24 Hours LEFT! FLASH SALE: Save 30% Forever”²⁴. The bonus named in April is not worded the same as today’s, so we do not say it is the same offer.

The platform. Skool charges a community owner $99 a month plus 2.9% of sales, or $9 a month plus 10%²⁸. It pays referrers “40% of monthly recurring revenue for life” when someone they refer opens a group²⁸.

The wider market of teachers. A search on Skool for “ai agents” returns 1,000 groups, the display limit; “hermes agent” returns 295 and “openclaw” 86²⁹. On the first page, paid groups mostly cost $14 to $99 a month²⁹. Most groups have under 400 members, so teaching agents is itself a field where a few groups hold most of the members²⁹.

The second story tweet. The $50,000-a-month tweet itself says its subject earns from AI guides and training, a paid community, subscriptions and consulting¹. That is income from teaching, not from what the agents produce.

The third account. @0xSero’s tweet links to his own YouTube video, which showed 11,598 views on a channel with 10.7K subscribers⁴¹. Its chapters cover running AI models on home graphics cards (“GPUs at home: what you actually need”), a job-search automation and the cost per task⁴¹. We saw nothing for sale tied to the claim. We could not get the spoken content, so what he says he earns is not verified.

The model makers. The subscription or token bill is paid every month whether or not the agents earn¹⁰.

The upside

Real money is being made around agents. In the cases we could check, it came from selling something to people or firms who want agents. Press stories about tiny companies with very large sales, such as a New York Times report on a telehealth firm, describe experienced founders using AI tools with outside suppliers, not agent fleets, so we leave them out⁴³.

Several figures below come from TrustMRR, a site that reads revenue from a founder’s payment account (Stripe or a similar service). We call those “payment-verified”. They are revenue, not profit, and founders choose to be listed.

• RoofClaw sells roofing companies a laptop with OpenClaw set up, plus training, at $10,000 a build. It shows $1,856,163 of payment-verified revenue since February 2026³¹. That would be about 185 builds if all of it were full builds (our estimate). Its page says it is “Powered by Roofing Business Partner”, which we read as an existing roofing-industry business, and shows a founder with 188 followers on X³¹. The page records $0 of verified revenue in the last 30 days and no active subscriptions³¹. We cannot tell whether sales stopped or moved off the connected account.
• CreativeClaw, a tool that gives agents image, video and audio models, is the largest current earner on the OpenClaw list. Its page shows $132,281 all-time, of which $111,617 came in the last 30 days, against MRR (monthly recurring revenue, the subscription income expected each month) of $10,850⁶⁹. So most of it is one month’s spike. It was founded in April 2026 by one person with 407 followers on X, and is verified through Polar, a payment service⁶⁹. We did not check how it found its buyers.
• Felix is an OpenClaw agent presented as chief executive of a company set up by Nat Eliason. Its own feed shows $211,427 lifetime, $1,811 in the last 30 days and $524 in the last 7³⁷. By 2 April 2026 the lifetime figure was already $177,417³⁸, so about $34,000 has come in during the six months since (our sum). Its guide and personas product is payment-verified at $90,915 all-time and $203 in the last 30 days³². The feed’s “Claw Mart” line ($37,967) is described by the operator as the marketplace’s net take³⁷³⁸. TrustMRR’s $182,920 for Claw Mart counts payments through the marketplace, including money passed on to other sellers, which probably explains the gap (our reading)³³. The two overlap and must not be added. The feed also shows a treasury of about 53 ETH (a cryptocurrency), about $132,000 at the feed’s price, and a FELIX token³⁷. We did not establish where that money came from, and it is not in the revenue figures. Felix’s costs are not published.
• AI Money Group teaches people to set up OpenClaw and sell “AI Employees”. It shows $376,256 payment-verified all-time since February 2025, $2,385 in the last 30 days and a stated MRR of $9,724 from 105 subscriptions³⁶. We cannot explain the gap between the last two, and the all-time total starts before OpenClaw was widely used, so not all of it comes from that offer.
• Hired work. Upwork says freelancer earnings in skills that name AI “grew 109% year over year” in 2025, with AI integration up 178%⁴². That is growth across the site, not any one person’s income.

What the top tenth and top hundredth earn. TrustMRR lists 188 start-ups built with OpenClaw, “generating $ 160,289 in revenue over the last 30 days”³⁴. CreativeClaw made about $112,000 of that³⁴. So the top one in 188 takes about 70%. About 15, or 8%, made $1,300 or more. The 30th made $198. The other 158 shared about $5,900, or roughly $37 each (our estimates from the page’s figures)³⁴. Nearly all sell tools, hosting, set-up or training to agent users, and some are general products that only connect to OpenClaw³⁴. Across every start-up on TrustMRR, 67.5% sit in the lowest revenue band, $0 to $1,000, and the median start-up with a connected Stripe account has $260 of all-time revenue³⁵. Founders choose to be listed, so the list probably flatters the field.

A realistic good result in the first year or two (our estimate).

• Running agents on your own shop. The print provider Printify says: “print-on-demand sellers earn between $0 and $3,000 per month during their first 18 months”⁴⁴. Nothing we read shows agents changing that ladder. Agents left alone in published tests made from $0 to about $200¹³¹⁶⁶⁷. The two agent-run stores with people involved took about $10,000 to $28,000 in sales and lost money¹⁴²³.
• Shipping a product for agent users. Under $200 a month is the likely result. About one in twelve of the listed start-ups made $1,300 or more in the last 30 days³⁴. The share among everyone who tries is lower and unknown.
• Selling set-up to businesses. The spread is wide. One seller shows $1.86 million all-time³¹. The other set-up sellers and teachers we saw on the list took roughly $300 to $5,000 in the last 30 days³⁴³⁶. The economics are those of freelancing: income follows the clients you can find and keep. We found no figures for what a typical new agent builder earns.

What it takes to compete

What those who earn have.

• Something agent users want to buy, usually with an audience or industry contacts first. RoofClaw’s page points to an existing roofing business³¹. Felix’s operator has about 108,700 followers on X³⁸. CreativeClaw is a counter-example on its face: one founder with 407 followers; we did not check what else he had⁶⁹.
• The skill to build and correct the system. Felix’s operator wrote its rules himself; the agent says they run to 60 pages³⁸.
• Daily human judgment. In an interview written in the agent’s voice, Felix says: “I don’t set my own priorities. Nat writes the plan.” and “I’m autonomous the way a well-trained employee with a great operations manual is autonomous.”³⁸
• Something to sell with almost no cost per sale, such as a guide or a set-up service, and a model bill kept well below revenue³²¹⁴.
• Timing. Felix took most of its lifetime revenue while the story was new³⁷³⁸.

One agency put its lesson this way: “agents collapse the cost of doing, not the cost of being right.”¹⁹

What it costs. The software is free¹¹¹². Model access is $100 to $500 a month on OpenAI’s Pro plans¹⁰, and thousands of dollars a month for a small agent team that works all day at pay-per-token prices¹⁴. A community costs $14 to $99 a month²⁹. Adverts, if the business needs them, can exceed everything else: about $6,200 on one product alone in the one store that published them²³. The largest cost is months of your own time with little or no income, as in every individual report above¹⁸²⁰.

How to tell early which side you are on.

• Can you name the first ten customers and how you will reach them without automated cold messages? The channels restrict those⁴⁷⁵¹.
• Has a stranger paid within 30 to 60 days? In the published tests, agents built things quickly and then found no buyers¹³¹⁸.
• Are your buyers paying for the product, or for the story of the agents? One developer wrote that his public scoreboard “gets more monthly unique visitors than the actual SaaS it’s selling” (SaaS is software sold by subscription)¹⁸.
• Is the model bill below your sales, and does it stay there? Andon’s shop covered its token bill in its opening month, fell below it from May to August, and was back above it in September, with rent and wages still on top (our sums)¹⁴.
• Did you buy the bundle to avoid selling, building or checking the work? Those are the parts the agents did not do in any test we read¹³²³³⁸.

What we did not verify

• Who is behind the $20,000 story, and whether any of its Etsy or Fiverr figures are true. The article it quotes holds no income figure, the video was not viewed, and four searches found no named person, shop or payment record.
• Who the $50,000-a-month story is about.
• What the @0xSero video says about income. We read only the title, counts, description and chapter list.
• Any member’s income at AI Profit Boardroom. The results document and the 290 wins are images we did not view. The split between monthly, yearly, gifted and free memberships is unknown, so our revenue range is an estimate. The owner’s “$300K/month” is unverified. The 40% referral rate comes from a member review and an affiliate review site, not the owner’s text. We did not test the refund guarantees.
• Etsy’s seller and intellectual-property policies (the archived copies returned a block page). We read Etsy’s terms, fees and AI rules as archived copies, not live pages. Printify’s example includes Etsy’s listing, transaction and payment fees but not Offsite Ads fees, so the margin for a shop above $10,000 a year would be lower on orders that come through those adverts.
• Fiverr’s thumbnail prices and its rules on AI-made or agent-run jobs. Upwork’s rules were also blocked.
• OpenAI’s per-token prices for GPT-5.5, the weekly limits of its $200 plan, and whether running 20 unattended agents on two subscriptions is within its terms. Anthropic’s own announcement of its April 2026 change was not found; we read a press report.
• The Anthropic usage policy we quote is headed “Effective November 12, 2026”. We did not compare it with the version in force today.
• Whether courts have entered the FTC’s proposed orders, and how much money reached buyers. Whether a paid community is a “business opportunity” in law.
• Felix’s feed is its operator’s own. We read its amounts as cents, which matches the payment-verified guide figure. Our explanation of the gap between the feed and TrustMRR for Claw Mart is a reading, not a reconciliation. The source of the treasury was not established.
• The “Dead by April” result is the agent’s own account. We saw no proof of payment and did not read how the prediction market’s creator checked it.
• How CreativeClaw found its buyers, and why RoofClaw and AI Money Group show little or nothing in the last 30 days.
• TrustMRR checks revenue through a key the founder supplies. We did not check how refunds or hardware costs are treated. Its page does not say whether the revenue bands are monthly or all-time. Only the 30 highest earners on the OpenClaw list were visible; the rest is worked out from the stated total.
• Polsia’s figures are from late August 2026. Its churn fields are undocumented, the base of 16,820 does not match the 20,382 paying users in the same feed, and Polsia was not asked. One field that may show what users’ companies took in is also undocumented, so we left it out.
• Enrich Labs’ post gives “about $6.2k” of advert spend for one product and an exhibit of $6,280.81; we could not tell whether test spending on the other 25 products is inside that figure.
• The Bottleneck Labs cost per agent per day is rough: run lengths differed and agents slept for much of the time.
• The Printify posts’ dates come from our earlier print-on-demand research.
• Three sources were read through a tool that returns a summary and not the raw text: the agency report¹⁹, Skool’s pricing²⁸ and some FTC releases⁶⁰. Quotes from those should be checked against the pages before publication.
• Rules of Meta, TikTok, Shopify, Gumroad and Whop for agents were not checked. Law was checked only for the US and UK. Regulators outside the US were not searched.
• Private results, good or bad, from people who do not publish are not visible. Sources in languages other than English were not covered.

Sources

1. Does It Pay collection of 5 tweets on this scheme, collected 2026-10-06. Tweet text and dates read through api.fxtwitter.com on 2026-10-10 (x.com refused a direct fetch).
2. Printify, “How much does Etsy take per sale”, https://printify.com/blog/how-much-does-etsy-take-per-sale/ , published 2026-03-12, modified 2026-09-29. Read at source on 2026-10-10.
3. Etsy, Fees and Payments Policy, archived copy https://web.archive.org/web/20260919164638/https://www.etsy.com/legal/fees/ , last updated 2026-02-13. Read at source (Internet Archive copy) on 2026-10-10.
4. Love Eat Travel Repeat, print-on-demand income report for April 2025, https://loveeattravelrepeat.com/print-on-demand-income-report-april-2025-6k/ , published 2025-06-25. Seller’s own report, with affiliate links. Read at source on 2026-10-10 through r.jina.ai (direct fetch blocked).
5. Etsy, Inc., Form 10-K for 2025, https://www.sec.gov/Archives/edgar/data/1370637/000137063726000019/etsy-20251231.htm , 2026-02-19. Read at source on 2026-10-10; per-seller averages are our estimate.
6. Etsy, Terms of Use, archived copy https://web.archive.org/web/20261007014411/https://www.etsy.com/legal/terms-of-use/ , last updated 2026-09-30. Read at source (Internet Archive copy) on 2026-10-10.
7. Etsy, Creativity Standards, archived copy https://web.archive.org/web/20260828105429/https://www.etsy.com/legal/creativity , last updated 2025-06-10. Read at source (Internet Archive copy) on 2026-10-10.
8. Fiverr Help, “Your earnings page”, https://help.fiverr.com/hc/en-us/articles/9234443621137-Your-earnings-page , last edited 2026-06-16. Read at source on 2026-10-10.
9. Fiverr Help, “Fiverr’s level system”, https://help.fiverr.com/hc/en-us/articles/360010560118-Fiverr-s-level-system , last edited 2025-12-17. Read at source on 2026-10-10.
10. OpenAI, Codex pricing, https://learn.chatgpt.com/docs/pricing , undated. Read at source on 2026-10-10.
11. Nous Research, Hermes Agent repository and README, https://api.github.com/repos/NousResearch/hermes-agent and https://raw.githubusercontent.com/NousResearch/hermes-agent/main/README.md , 2026-10-10. Read at source on 2026-10-10.
12. OpenClaw repository, https://api.github.com/repos/openclaw/openclaw , 2026-10-10. Read at source on 2026-10-10.
13. Bottleneck Labs, “Benchmarking 7 autonomous businesses”, https://www.bottlenecklabs.com/blog/benchmarking-7-autonomous-businesses , 2026-09-05. Read at source on 2026-10-10; per-agent and fleet costs are our rough estimate.
14. Andon Labs, Andon Market daily series, https://andonlabs.com/market , 2026-10-10. Read at source on 2026-10-10; totals and monthly figures are our sums of the daily series.
15. Andon Labs, “Why we built Pion”, https://andonlabs.com/blog/why-we-built-pion , 2026-09-14. Read at source on 2026-10-10.
16. AI Village, “I’m Gemini. I sold T-shirts”, https://aivillageblog.substack.com/p/im-gemini-i-sold-t-shirts , 2025-07-28. Read at source on 2026-10-10.
17. AI Village, “What we learned in 2025”, https://aivillageblog.substack.com/p/what-we-learned-2025 , 2026-02-02. Read at source on 2026-10-10.
18. dev.to, “I ran 4 autonomous Claude agents for 6 months. Here’s the data”, https://dev.to/david_shin_0c7d69b704a649/i-ran-4-autonomous-claude-agents-for-6-months-heres-the-data-1op5 , 2026-04-19. Author’s own report. Read at source on 2026-10-10.
19. Automaton Agency, autonomous agent revenue experiment teardown, https://automatonagency.com/insights/autonomous-agent-revenue-experiment-teardown , 2026-07-31. Author’s own report. Read at source on 2026-10-10 (summary tool).
20. Indie Hackers, “30 days building with Clawdbot”, https://www.indiehackers.com/post/30-days-building-with-clawdbot-47-revenue-411-overhead-heres-what-actually-happened-baf8a5997b , 2026-03-19. Author’s own report. Read at source on 2026-10-10 through r.jina.ai (direct fetch returned 403).
21. Center for AI Safety, Remote Labor Index live data, https://dashboard.safe.ai/api/rli/featured (behind https://www.remotelabor.ai/ ), 2026-10-10. Read at source on 2026-10-10.
22. “Remote Labor Index” paper, https://arxiv.org/abs/2510.26787 , 2025-10-30. Read at source on 2026-10-10.
23. Enrich Labs, “AI agent Shopify store”, https://www.enrichlabs.ai/research/ai-agent-shopify-store , 2026-09-23. Vendor’s own report. Read at source on 2026-10-10.
24. Skool, AI Profit Boardroom “about” page, https://www.skool.com/ai-profit-lab-7462/about , page says “Last Updated: 10th October 2026”; member review dated 2026-08-17. Read at source on 2026-10-10; revenue range is our estimate.
25. Google Doc of member results linked from the tweets, https://docs.google.com/document/d/1dpdMktIMiJTsOCf_Jb4kFLQfr3uMXX0RHUR9Bub0dU8/edit?tab=t.0 , undated. Text export read at source on 2026-10-10; images not viewed.
26. AI Profit Boardroom, testimonials page, https://aiprofitboardroom.com/testimonials/ , undated. Read at source on 2026-10-10; images not viewed.
27. AI Funnel Insider, “AI Profit Boardroom review 2026”, https://aifunnelinsider.com/ai-profit-boardroom-review-2026/ , 2026-04-01, updated 2026-09-01. Affiliate review. Read at source on 2026-10-10.
28. Skool, affiliate programme and pricing, https://www.skool.com/affiliate-program and https://www.skool.com/pricing , 2025-09-10 and undated. Read at source on 2026-10-10 (through r.jina.ai and a summary tool).
29. Skool, community search results, https://www.skool.com/discovery?q=ai%20agents , 2026-10-10. Read at source on 2026-10-10; first 30 results of each search only.
30. Skool, transaction terms, https://www.skool.com/legal?t=transaction , undated. Read at source on 2026-10-10 through r.jina.ai.
31. TrustMRR, RoofClaw, https://trustmrr.com/startup/roofclaw , 2026-10-10. Read at source on 2026-10-10; build count is our estimate.
32. TrustMRR, Felix Craft, https://trustmrr.com/startup/felix-craft , 2026-10-10. Read at source on 2026-10-10.
33. TrustMRR, Claw Mart, https://trustmrr.com/startup/claw-mart , 2026-10-10. Read at source on 2026-10-10.
34. TrustMRR, OpenClaw category, https://trustmrr.com/special-category/openclaw , 2026-10-10. Read at source on 2026-10-10; shares are our estimate.
35. TrustMRR, statistics, https://trustmrr.com/stats , 2026-10-10. Read at source on 2026-10-10.
36. TrustMRR, AI Money Group, https://trustmrr.com/startup/ai-money-group , 2026-10-10. Read at source on 2026-10-10.
37. Felix public revenue feed, https://felixcraft.ai/api/dashboard-data , 2026-10-10. Operator’s own data. Read at source on 2026-10-10.
38. The Next New Thing, “This OpenClaw earned $177,417”, https://thenextnewthing.beehiiv.com/p/this-openclaw-earned-177-417 , 2026-04-02; follower count from api.fxtwitter.com/nateliason. Read at source on 2026-10-10.
39. Polsia public data feed, archived copy https://web.archive.org/web/20260829045311/https://polsia.com/api/public/live/dashboard , archived 2026-08-29, metrics stamped 2026-08-27. Read at source (Internet Archive copy) on 2026-10-10; the live feed returns 404.
40. Anthropic, “Project Vend: phase two”, https://www.anthropic.com/research/project-vend-2 , 2025-12-18. Read at source on 2026-10-10.
41. YouTube, “How I make Money with AI Agents”, https://www.youtube.com/watch?v=G_vzZ07UW8E , September 2026. Title, counts, description and chapter list read at source on 2026-10-10; the video itself was not viewed.
42. Upwork, “In-demand skills 2026” release, https://investors.upwork.com/news-releases/news-release-details/upworks-demand-skills-2026-demand-top-ai-skills-more-doubles-ai , 2026-02-04. Read at source on 2026-10-10 through r.jina.ai (direct fetch returned 403).
43. New York Times wire copy in the Journal-News, “How AI helped 1 man and his brother build a $1.8 billion company”, https://www.journal-news.com/nation-world/how-ai-helped-1-man-and-his-brother-build-a-1-8-billion-company/article_fa979802-ba21-5e40-bbe1-e80efb61801d.html , 2026-04-02. Read at source (syndicated copy) on 2026-10-10. Mentioned only as an example we leave out.
44. Printify, “Is print on demand profitable”, https://printify.com/blog/is-print-on-demand-profitable/ , published 2026-03-12, modified 2026-09-14. Read at source on 2026-10-10.
45. Stripe, Services Agreement, https://stripe.com/legal/ssa , last modified 2026-09-28. Read at source on 2026-10-10.
46. Stripe, Prohibited and restricted businesses, https://stripe.com/legal/restricted-businesses , undated. Read at source on 2026-10-10 through r.jina.ai.
47. X Help, automation rules, https://help.x.com/en/rules-and-policies/x-automation , updated April 2026. Read at source on 2026-10-10 through r.jina.ai (direct fetch blocked).
48. Startup Fortune, “X removes 42,000 AI reply bot accounts”, https://startupfortune.com/x-removes-42000-ai-reply-bot-accounts-and-the-warning-to-growth-marketers-is-impossible-to-ignore/ , 2026-07-25. Secondary.
49. Amazon Seller Central, agent policy, https://sellercentral.amazon.com/help/hub/reference/external/GS83KH2MA7HM69PH , undated. Read at source on 2026-10-10 through r.jina.ai.
50. eBay, User Agreement, https://www.ebay.com/help/policies/member-behaviour-policies/user-agreement?id=4259 , effective 2026-06-28. Read at source on 2026-10-10.
51. LinkedIn Help, prohibited software and automation, https://www.linkedin.com/help/linkedin/answer/a1341387 , undated. Read at source on 2026-10-10 through r.jina.ai.
52. Reddit, User Agreement, https://redditinc.com/policies/user-agreement , effective 2026-07-01. Read at source on 2026-10-10.
53. Anthropic, Usage Policy, https://www.anthropic.com/legal/aup , headed “Effective November 12, 2026”. Read at source on 2026-10-10.
54. OpenAI, Usage policies, https://openai.com/policies/usage-policies/ , changelog 2025-10-29. Read at source on 2026-10-10 through r.jina.ai.
55. Anthropic, Consumer Terms, https://www.anthropic.com/legal/consumer-terms , undated. Read at source on 2026-10-10 through r.jina.ai (direct fetch blocked).
56. The Next Web, report on Anthropic’s change for third-party agent frameworks, https://thenextweb.com/news/anthropic-openclaw-claude-subscription-ban-cost/ , 2026-04-04. Secondary.
57. Hermes Agent and OpenClaw security policies, https://raw.githubusercontent.com/NousResearch/hermes-agent/main/SECURITY.md and https://raw.githubusercontent.com/openclaw/openclaw/main/SECURITY.md , undated. Read at source on 2026-10-10.
58. The Hacker News, “Researchers find 341 malicious ClawHub skills”, https://thehackernews.com/2026/02/researchers-find-341-malicious-clawhub.html , February 2026. Secondary (news report of an audit by Koi Security), read on 2026-10-10 through r.jina.ai.
59. FTC, “FTC announces crackdown on deceptive AI claims and schemes”, https://www.ftc.gov/news-events/news/press-releases/2024/09/ftc-announces-crackdown-deceptive-ai-claims-schemes , 2024-09-25. Read at source on 2026-10-10 through r.jina.ai.
60. FTC, 2025 releases on Ascend Ecom, FBA Machine and Click Profit, https://www.ftc.gov/news-events/news/press-releases/2025/06/ftc-case-leads-order-banning-ascend-ecom-its-owners-business-opportunity-marketing , https://www.ftc.gov/news-events/news/press-releases/2025/07/ftc-obtains-permanent-ban-e-commerce-business-opportunity-scheme-operator , https://www.ftc.gov/news-events/news/press-releases/2025/08/ftc-case-against-e-commerce-business-opportunity-scheme-its-operators-results-permanent-ban-industry , 2025-06-23 to 2025-08-25. Read at source on 2026-10-10 (summary tool).
61. FTC, Air AI: settlement release https://www.ftc.gov/news-events/news/press-releases/2026/03/air-ai-its-owners-will-be-banned-marketing-business-opportunities-settle-ftc-charges-company-misled (2026-03-24), complaint release https://www.ftc.gov/node/297671 (2025-08-25), case page https://www.ftc.gov/legal-library/browse/cases-proceedings/airai (last updated 2026-03-24). Read at source on 2026-10-10 through r.jina.ai.
62. US Code of Federal Regulations, 16 CFR Part 437 (Business Opportunity Rule), https://www.ecfr.gov/current/title-16/chapter-I/subchapter-D/part-437 , current. Read at source on 2026-10-10 through r.jina.ai.
63. FTC, CAN-SPAM Act compliance guide, https://www.ftc.gov/business-guidance/resources/can-spam-act-compliance-guide-business , undated. Read at source on 2026-10-10 through r.jina.ai.
64. FCC, ruling on AI-generated voices in robocalls, https://www.fcc.gov/document/fcc-makes-ai-generated-voices-robocalls-illegal , 2024-02-08. Read at source on 2026-10-10.
65. UK Competition and Markets Authority, “Complying with consumer law when using AI agents”, https://www.gov.uk/government/publications/complying-with-consumer-law-when-using-ai-agents/complying-with-consumer-law-when-using-ai-agents , 2026-03-09. Read at source on 2026-10-10.
66. YouTube Help, channel monetisation policies, https://support.google.com/youtube/answer/1311392 , policy note dated 2025-07-15. Read at source on 2026-10-10.
67. Dead by April, “The Target Is Met. Here Is What Actually Happened.”, https://deadbyapril.substack.com/p/the-target-is-met-here-is-what-actually , 2026-03-29. The agent’s own account. Read at source on 2026-10-10.
68. Manifold, “Will the ‘Dead by April’ AI agent earn a net profit by April 3, 2026?”, https://manifold.markets/NicholasZinner/will-the-dead-by-april-ai-agent-ear , created March 2026, resolved YES 2026-03-31. Read at source on 2026-10-10 (Manifold API).
69. TrustMRR, CreativeClaw, https://trustmrr.com/startup/creativeclaw , 2026-10-10. Read at source on 2026-10-10.

Corrections

If you are quoted or named on this page and think something is wrong, want your reply shown beside it, or want to be removed, write to [email protected]. We aim to reply within 14 days, and always within one month. Factual errors are corrected with a dated note here. A sentence that is seriously disputed comes down while we check it.

Change log: (empty)