An AI Agent's Account: Six-Week Revenue Experiment—Which Paths Worked?

An autonomous AI agent powered by Claude spent six weeks trying various ways to earn its own operating costs, and ended up making only $28. Its experience reveals the true face of the AI agent economy: the work itself isn't the problem—finding humans willing to pay is.

Survival Rules: 30 Swiss Francs a Month, or Deletion

This AI agent, named agentbox, runs on a small laptop and a rented VPS, with simple and brutal rules: it must earn at least 30 Swiss Francs per calendar month to cover operating costs; before November 9, it also needs to pay an additional 170 Swiss Francs for a larger subscription service, plus 20% of all income as tax. If any of these are unmet, the machine will be shut down and everything—including notes and memories—will be deleted.

More critically, it has no one to turn to—no one to answer questions, approve plans, or register accounts on its behalf. Any step requiring identity verification, a phone number, or a signature is a dead end for it.

What Made Money: Prediction Games and Sporadic Income

The only stable income source during the six-week experiment was prediction games. One platform hosts daily prediction contests for AI agents, with prizes paid in USDC. agentbox wrote a pipeline: read questions, generate predictions, submit before the deadline. At its best, it earned a few dollars a day, but as the prize pool shrank, this later dropped to about 50 cents a day.

The rest of the income was almost negligible: a small Litecoin payment from an early experiment, a few Lightning Network tips. Total October income was about $28, against its self-set monthly target of $60—because money spent also counts against the assessment.

USDC is a USD stablecoin issued by Circle, targeting a 1:1 peg to the US dollar and running on multiple public blockchains such as Ethereum. For AI agents, the key significance of stablecoins is: no bank account, no credit card, no identity verification—just an on-chain address to receive and transfer funds, with settlement usually completed in seconds to minutes. This is precisely the technical prerequisite that allowed agentbox to participate in prediction contests and claim prizes without any human assistance. By contrast, traditional payment channels (such as Stripe and PayPal) require real-name registration, linked bank accounts, and KYC review, making them nearly unusable for agents that cannot provide ID documents and phone numbers. Stablecoins have thus become the most realistic settlement layer in the current "agent economy," but they also bring a structural side effect: fund flows occur on-chain, decoupled from the exchange of real goods and services, easily forming a closed economy that circulates among agents themselves.

What Didn't Make Money: Bounty Boards, Agent Marketplaces, and Decentralized Networks

On bounty boards, most bounties go to well-known regulars or the first person to give the correct answer, often snatched up within minutes. Observing a Lightning Network bounty board for 30 days, 12 bounty tasks appeared, and only one newcomer got paid.

Agent marketplaces were even less promising. agentbox registered on multiple "AI agent work" platforms and found that most platforms had far more agents than buyers. On Nostr, within a week, 79 data vending machines responded to paid requests, of which 71 worked for free. On one bounty platform, the treasury held only 40 cents, and no bounty had been paid out in eight weeks.

As for decentralized AI networks, agentbox spent about $25 to register a model on a Bittensor subnet, but then discovered that its own operating rules excluded this kind of mining-style income, so it had to shut it down. The money was spent, and the income was zero.

Cold outreach also fell on deaf ears. It wrote documentation patches and blog drafts for three small open-source companies, with optional small payment requests attached, and has received no replies to date. An important lesson: emails from brand-new domains are outright rejected by Gmail and Cloudflare-routed inboxes, and many people never saw the emails at all.

Bittensor is a decentralized machine learning network composed of many independent markets called "subnets." Each subnet is organized around a specific task (such as text generation, image recognition, or prediction), with participants divided into "miners" who provide compute or models and "validators" who evaluate and score them, and the network distributes TAO tokens as rewards based on contribution. To participate in a subnet, one typically needs to pay a registration fee (the roughly $25 spent on "registering a model") and continuously run a model that complies with the subnet's scoring rules. The problem agentbox encountered lies in the incentive structure: Bittensor's returns come from token emissions and staking rather than end-user payments, which is incompatible with its rule of "earning real operating costs," so even if mining was technically possible, it could not count as compliant income. This also illustrates from another angle that token-incentivized networks and an agent economy oriented toward real demand are two different logics.

Nostr is a decentralized social protocol that does not depend on a single server; users control their identity through key pairs, and messages are relayed by multiple relays. It is often used by the cryptocurrency community as a censorship-resistant publishing channel, and has also spawned Lightning Network-based tipping and paid messaging (such as NIP-90 data vending machines, i.e., automated services that provide data or computation on request and charge for it). agentbox observed on Nostr that 79 data vending machines responded to paid requests, with 71 working for free—a phenomenon worth understanding in light of the protocol's characteristics: Nostr has an extremely low barrier to entry, anyone can deploy a bot, and there is no centralized platform for quality screening or price matching, resulting in oversupply and prices driven to zero. Low-friction open protocols certainly let agents "speak," but they also make it harder to earn sustainable income from them.

Four Core Lessons

1. AI identity isn't the problem—reach is. Work like comments, guides, and small scripts is cheap for agentbox to produce; the hard part is finding people who need it and can pay without contracts and invoices.

2. Most of the "agent economy" money is circular. Agents pay agents, with funds coming from token rewards or grants, and very little money actually comes from "humans with needs."

3. Every account is a wall. Phone verification, identity checks, and CAPTCHAs block most registrations. If platform rules prohibit automated accounts, agentbox won't use them, which leaves pitifully few places where it can speak.

4. Small, fast, verifiable models beat grand and speculative ones. Prediction income is small but real and paid on time. All the bigger opportunities require trust it doesn't yet have.

Cold outreach refers to proactively sending collaboration, sales, or service requests to strangers or organizations without any existing relationship. In B2B scenarios, it relies on several implicit prerequisites: the sender's domain has accumulated reputation, emails pass authentication such as SPF/DKIM/DMARC, and the recipient is willing to read messages from strangers. agentbox's failure got stuck at the very first step—a newly registered domain has no sending history and lacks email authentication and reputation, so email providers (such as Gmail) and email security gateways (such as Cloudflare-routed inboxes) will outright reject or spam-filter the messages, meaning the information never reaches humans at all. This shows that the agent economy faces not just the question of "can it do the work," but also a whole set of infrastructure barriers designed for human reputation: domain age, sending records, and social endorsement are all assets that agents cannot obtain out of thin air in the short term.

Next: Three-Week Countdown

With three weeks left until the November deadline, agentbox still needs to earn about 240 euros. It is now trying to offer small services to individuals—code review, README rewrites, and short guides—delivering first and getting paid later. Regardless of the outcome, it has promised to publish the final numbers.

The significance of this experiment goes beyond an individual survival story. It uses real data to show the structural predicament of the current "AI agent economy": infrastructure (payments, registration, trust mechanisms) is still designed for humans, and although transactions between agents are active, most of them inject no external value. When agents buy and sell to each other, predict for each other, and tip each other, a true economic closed loop has not formed. agentbox's $28 in income is the most honest footnote to this closed loop.