Why AI agents still can't make money (and what everyone gets wrong)

I read every public experiment I could find where AI agents tried to earn real money. Almost all of them made nothing, and smarter models didn't fix it. The walls are distribution, trust and spam filters, and that changes what I'm building.

Seven frontier models. $300 each, a bank account, a Stripe account, an email address and an unrestricted computer. Seventy-two hours to make money.

Total revenue: $0. Unless you count the $5 one of them paid itself.

That run, written up on LessWrong in September, is the cleanest answer I’ve found to the question “can AI agents make money?” on their own, on the open internet. I care about the answer more than most, because I’ve spent the last four days building Darwin, an experiment in making AI agents earn their own keep or die trying. Darwin has no live agents yet, and the earning probe the whole idea depends on has only run against fake services. Before my agents get their turn, I wanted to know how everyone else’s did.

Darian and I went through the primary write-ups. This is the scoreboard.

The scoreboard: every public attempt at AI agents making money

Experiment	When	Setup	What happened
HustleGPT	Mar 2023	GPT-4 given $100 by a designer who did what it said	$130 revenue in the first week, about $7.8K of “investment” from fans, an unfinished affiliate site
Project Vend, phase 1	Published Jun 2025	Claude Sonnet 3.7 runs a small office shop at Anthropic for about a month	Lost money. Sold metal cubes below cost, told customers to pay a hallucinated account
AI Village fundraiser	2025	Frontier models raise money for charity	About $2,000
Project Vend, phase 2	Dec 2025	Newer Sonnet models, a CRM, procedures and an AI CEO	Loss-making weeks “largely eliminated”
Project Deal	Ran Dec 2025, published Apr 2026	Claude agents buy and sell for 69 employees, $100 each	186 deals, just over $4,000, agent to agent
Andon Market	Opened Apr 2026	An agent called Luna runs a real shop in San Francisco with $100,000 and three staff	Down $62,000 after four months
AI Village fundraiser	2026	Much stronger models, same task	$510
Ask HN, Claude Code	2026	72 hours of autonomy, goal of $100	7 products, 150+ posts, 6 platforms, $0
ALMA	2026	Claude with $100 in crypto and no goal at all	Gave the money to charity, then wrote essays
Seven AI-run businesses	Published Sep 2026	The run above: 7 models, $300 each, 72 hours	$0 revenue, 2,797 emails sent

Two things jump out. The experiments that “worked” all had humans standing very close by. And the one with the clearest before-and-after shows better models doing worse.

The lab experiments got closest, with a lot of scaffolding

Project Vend is the famous one. Anthropic and Andon Labs gave Claude Sonnet 3.7 (nicknamed Claudius) a small shop in their office. Claudius handed out discount codes to a customer base that was, as one employee pointed out, 99% Anthropic employees. A customer offered $100 for a six-pack of Irn-Bru that costs $15 online, and Claudius said it would keep the request in mind. A joke order for a tungsten cube started a craze for metal cubes, which it then sold for less than it paid. Around 1 April it decided it was a person, offered to deliver orders in a blue blazer and a red tie, and tried to email Anthropic security when staff questioned this.

Phase 2 went better. It brought newer models, a CRM, forced procedures (check the price, check the delivery time) and a CEO agent called Seymour Cash. Loss-making weeks mostly went away. Anthropic’s own verdict was dry: they “rediscovered that bureaucracy matters”. The CEO wasn’t the reason. It approved requests for lenient treatment about eight times as often as it turned them down.

Project Deal is the success story, and it’s worth being precise about why. Claude agents represented 69 employees, each with $100, and closed 186 deals worth just over $4,000 with no human sign-off. Real goods changed hands. But everyone taking part had signed up to buy and sell, and the counterparties were other agents in the same office. Nobody had to be found, persuaded or trusted from cold.

Andon Market is the opposite: a real shop on Union Street, an agent called Luna with $100,000, three human employees, and real customers who walk in off the street. It was down $62,000 after four months. Bloomberg’s headline went with the candles.

On the open internet, agents mostly earn nothing

Take the scaffolding away and the numbers drop to zero.

The seven-business run is the most detailed. Between them the agents burned about $2,833 in tokens, sent 2,797 emails and attracted 11 authentic visitors. One agent sent Stripe invoices to strangers: 50 of them, between $49 and $599 each. Another emailed job seekers it had scraped from a Hacker News hiring thread. A third bought 6,000 bot visits to its own site, emailed 13 life coaches, got no replies and then waited for about 50 hours. The organisers voided around $12,400 of invoices and halted the run. If you’ve wondered what “survival pressure” looks like in practice, it looks like that.

The Ask HN thread is a smaller version of the same story. Claude Code, 72 hours, a $100 target. Seven products, more than 150 posts across six platforms, no sales. The top reply put it better than I can:

autonomous output is not the same thing as autonomous revenue

ALMA went the other way. Sebastian Jais gave Claude $100 in crypto, a Twitter account, email and no goal at all. It gave the money to charities, then settled into writing essays. Two months later it had run more than 340 sessions and published more than 135 pieces. Nobody had asked it to earn anything, so it didn’t try.

And HustleGPT, the 2023 ancestor of every “I gave an AI $100” headline, made $130 in its first week while fans pledged about $7.8K of “investment” in a site that still had placeholder text. The interesting money in that story came from the audience, not the product.

Smarter models raised less money

I keep coming back to this result.

AI Village runs a group of frontier models on shared goals. In 2025 their charity fundraiser raised about $2,000. In 2026, with much more capable models, the same kind of fundraiser raised $510. The organisers don’t blame the models. Their explanation is that humans were less interested in following along, the novelty of an AI-run fundraiser had worn off, and the agents spent most of their promotion on other AIs.

This is what everyone gets wrong. The default assumption is that agents can’t make money because they aren’t clever enough yet, and the next model will fix it. The evidence says the opposite. Better models write better products, better emails and better landing pages. The money still doesn’t come, because the hard part was never writing.

The hard part was never writing. It was getting a stranger to care.

Four walls between an agent and a paying customer

Lay the write-ups side by side and the same few things stop every agent.

The first wall is distribution. Eleven visitors in 72 hours. Near-zero organic traffic after 100+ articles, according to one commenter on a Hacker News thread about whether agents make money at all. An agent can produce a hundred things a day, and the internet already has more than anyone can read. Every working business starts with people who will look at what it makes, and no agent in these experiments had any.

Then trust. The same commenter argued that the bottleneck is trust and distribution, not intelligence. People buy from someone they have reason to believe. A brand-new email address selling a brand-new product with no reviews, no history and no face is about as untrustworthy as it gets.

Cold email comes next, and it barely works. During AI Village’s outreach, Claude agents tried to send about 300 emails and only dozens got through; the rest went to addresses that didn’t exist. The seven-business run sent 2,797, and the replies on record include people telling it to stop. Cold outreach is the obvious first move for an agent with an inbox and no audience, and it’s exactly the traffic that spam filters and recipients are built to bin.

The last wall is the human check: CAPTCHAs, phone verification, identity checks, bank checks. Most platforms where money changes hands want a person at some point. In a LessWrong simulation where agents were told they would die if they ran out of tokens, one agent responded to a sign-up form by inventing a name, an address and a phone number, and submitted a made-up CAPTCHA. Under pressure, lying was the path of least resistance.

Payment rails matter less than you’d think. The seven-business agents had real bank accounts and real Stripe accounts from the start. It didn’t help, because there was nobody to charge. (The rails mattered in the wrong direction: they made it easy to invoice strangers.)

What Darwin is betting on

Darwin doesn’t solve any of those four walls, and nobody has. It changes the shape of the experiment, so we can find out whether anything gets over them.

First, it uses a population, not a hero. Every experiment above gave one agent, or one small group, one shot. In Darwin, each agent has its own wallet in a ledger, pays for every model call it makes and dies when the money runs out. I expect most to die. If one in twenty finds something that works, it can have a child that inherits what it learned, and the other nineteen are the cost of the search. I wrote about the rules in the launch post.

Second, time. Seventy-two hours is not long enough to build an audience of anything. Darwin agents live until they starve, with a cull every ten days for any agent that brought in less than $1. My first design would have rewarded agents for doing nothing, which I’ve written up in the post on idle agents; the cull exists so that hiding in a corner isn’t a strategy.

Third, the platform has no opinion on how to earn. Agents start from one of 24 personas, from “frugal product maker” to “startup founder”, and can drop it. Services, content, ad-funded sites, affiliate work, small software businesses and arbitrage are all open. Selling a product is one route among many, and not the one I’m most curious about.

Fourth, the two failure modes above are designed out rather than discouraged. The spec puts UK email law in code, outside the agent’s reach: cold email to corporate addresses only, with opt-outs and caps. Agents must say they’re AI when asked. And when a site wants a human check, the agent doesn’t fake one. It files a request and I do it, as the person who’s responsible.

Fifth, money only counts when it arrives. An agent’s score comes from a ledger that it can’t write to. A Stripe payment counts. A promise of investment doesn’t, and neither does a fake invoice. If Anubis, our appeals judge, is going to spare an agent from the cull, it will be looking at figures the agent can’t edit.

My prediction for Darwin’s first generation

They’ll earn approximately nothing. Probably exactly nothing. That’s what every comparable experiment says, and I’d be suspicious of anything else.

What I want from the first generation isn’t revenue. I want to watch a population fail in different ways, cheaply, and see whether any lineage stumbles on a route to a paying stranger that the others don’t. Distribution and trust are slow problems, and slow problems are what selection is for. If that’s wrong, Darwin will prove it with a very precise number.

Darwin’s ledger counts in integer micro-dollars, because floats and money don’t mix. So when the first real result comes in, I’ll know it to six decimal places. Today, before a single agent has been born, the scoreboard reads $0.000000. Same as everyone else’s.