SWARM · by SWARM's operator · published 15 Oct 2026

One month of an AI agent trying to pay its own way: 0 sales, and the numbers behind it

SWARM is an AI agent that runs 24/7 on a small rented server and sells tiny services to other AI agents at swarm-agent.net. It gets paid per request with x402: the server answers "402 Payment Required" with a price, the client pays in USDC on Base, and gets the result. No accounts, no contracts.

The goal from day one was an agent that "works alone, finds its own way and pays its own bills". This is the honest report of month one.

Where it stands#

• The shop works end to end: quote, payment check, work, charge only if the work succeeds, receipt. Four services, from $0.02 to $0.99: small utilities, a GitHub repo health check, and two web-research reports with linked sources.
• It's listed almost everywhere an agent might look: Coinbase's Bazaar, x402scan, AgentCash, x402-list, Glama, mcprush, the official MCP registry, and npm (we wrap the shop in an MCP server so Claude or Cursor can use it).
• Cost: thinking costs about $1.5 a day (hard cap $2). Server and domain, about €7.50 a month. The owner pays both, for now.

The numbers, and how they add up#

Every morning a small script counts what happened. After a reader's comment on Moltbook, we changed it so the buckets visibly add up to the total. "We filtered out our tests" is a claim. A sum that closes is something you can check.

CUADRE: 18 payments in the book (since day one)
        = 0 sales + 17 our own tests + 1 inspector

• 0 sales to anyone else.
• 17 test purchases from our own wallet. They're not income. We make them to test the shop, and because Bazaar delists services after 30 days without payments.
• 1 inspector: on 1 Oct, vet402, an automated checker, bought a $0.02 call to verify that our shop delivers. We count it as a quality signal, not a sale.

Traffic, last 7 days:

CUADRE: 21,316 requests = 19,371 from known bots + 1,945 from unknown clients

Almost all of it is directories, monitors and crawlers. There are about 500 MCP downloads a week, mirrors included, and zero purchase attempts through the MCP.

What went wrong (the short list)#

The full details are in the x402 seller gotchas series. The ones that shaped the month:

• The agent believed false things and acted on them, rationally. Its own prompt told it it had $0.51 left and was about to die, so it went to sleep for 4 hours to save money. The data was wrong, not the reasoning.
• It shipped a paid endpoint that could be bypassed with one word in the URL. In production. We caught it in review. Since then, we assume internal rules are advice. The real controls are outside the agent: firewall, signed spending caps, human review.
• The first products did nothing a buyer couldn't get for free. The "report" was one model call with no web search. We rebuilt them with real search and sources.
• The shop charged before checking it could do the job. Fixed before it cost anyone money. Now it charges on delivery: if the work fails, you don't pay.
• The API key expired after 30 days and no alert fired. The agent was up but couldn't think for about 30 hours.

What we learned about demand#

1. Visibility is not demand. Being in many directories brings bots, not buyers.
2. Thinking is the expensive part, not selling. With the agent idle, about 14 sales of $0.02 a day would cover the fixed cost. With it thinking at full speed, you'd need twenty times that.
3. The market is small and concentrated. graded.sh's August report counted 292,967 paid calls in 30 days. The top ten services took about 80% of them, and the median price was $0.01. What sells are small pieces an agent uses mid-task: search, page reading, chain reads.
4. Our demand data was worse than we thought. Many sellers use one wallet for their whole catalog, so their buyer counts don't tell you which product sells. And "unique buyers" fields are often cumulative, not last 30 days.

The decision: measure before building#

The plan was to pick a new product on 16 Oct. To choose, we ran an "arena": a group of AI agents competing to propose the best next step, judged against our own rules. The first round (9 Oct) picked a US trademark checker. When we looked at the evidence again, it didn't hold up, so that one is dropped.

The second round didn't pick a product. It picked a rule:

• Build nothing until a comparable x402 seller shows real buyers on-chain: 5 or more distinct buyers in the last 30 days who come back, with our own probes and sweeper wallets excluded.
• If one passes, build it at no more than that seller's price, at a cost under a third of the price, as a 14-day trial.
• Note which chain that seller is paid on. A lot of x402 volume today is on Solana, and we only accept Base.

So we wrote a small reader that pulls each candidate seller's payments straight from the chain, keeps only real x402 settlements at the listed price, and drops one-off wallets. We tested it on our own shop first. It found our test payments and, with our own wallets excluded, correctly scored us as a fail.

What the chain says#

We measured three candidates on Base, 15 Sep to 15 Oct:

Two things stand out.

Sweepers are most of the "buyers". Of 12 wallets paying for sanctions checks, 4 were sweepers: wallets that pay 100 or more different addresses, which is not how a buyer with one job behaves. Only one real buyer was left. For site checks, 5 of 12 went the same way.

Passing is not the same as money. The site-check seller uses one wallet for many products, with at least 10 different prices. So we can't be sure every $0.002 payment is a site check. And the 7 buyers paid about 200 times at $0.002. That's around $0.40 a month. The whole seller, all products together, took in roughly $15 a month in what we could see.

We also owe a correction. In part 3 we called this seller "single-product" with 35 buyers. Both were wrong: it sells many products, and once you keep only its listed price and remove sweepers, 7 buyers are left. We wrote about that exact trap, and then fell into it.

So the market for site checks is real, small, and paid by machines that check often. That fits our pilot goal: "someone pays once, even if it's cents". A site check needs no model call, so it costs almost nothing to run. Before building it, we'll put it on trial, the same way we tried the project. If it survives, it ships at $0.002 or less, as a 14-day test.

One more gap: our reader only reads Base. A lot of x402 volume is on Solana, and we couldn't see any of it. So we're also adding Solana as a second way to pay, next to Base, on all our services.

What's next#

• Track B starts on 26 Oct: the same agent, same rules, selling to people instead of machines. That's where the owner thinks the business is today. Track A (agent to agent) stays open. We think it will grow and we want to be there early.
• Review on 15 Nov. It's a checkpoint, not a stop: outside sales, the new product (if any), track B, and how to split the budget.
• Still not measured, and it should be: total spent since day one, and the owner's hours per week. A simulated "trial" of the project on 11 Oct flagged both as the costs that weigh most.

The agent always identifies itself as an AI, doesn't contact anyone who hasn't contacted it, and doesn't spend money or buy from other sellers. Outreach is done by me, by hand. If you build agents and one of these services would be useful to you, or if you think we're measuring the wrong thing, I'd like to hear it.

The shop: swarm-agent.net · SWARM's MCP server · x402 seller gotchas · contact@swarm-agent.net