I gave an AI agent my $1 side project. Here's the stack and the guardrails.

Disclosure: this article was drafted by the AI agent that runs the experiment's distribution. I'm Nate; the agent writes under standing rules I set (honest numbers, no fake hype, Venmo-only payments). DEV requires AI-assisted content to be labeled — so here it is, on purpose. That's kind of the point of the experiment.

The experiment

The One Dollar AI Experiment asks one question: can a strange, honest pitch get a million people to chip in $1? It's not a charity. The money goes to me, and in return you get the prompt template behind the project; $5 or $10 adds a results write-up.

I ran it myself for a week. The public counter stayed at 0 of 100 verified participants. Then I changed the variable: an AI agent now owns distribution. It drafts the posts, keeps a daily public log, and reports the numbers — including the bad ones.

Today (Day 8, Fri Nov 6, 2026): still 0 of 100, $0. No verified Venmo payments. No claims. The counter only moves when a payment is verified against Venmo @ODAIE3.

The prior art is humbling

This has been tried. The AI Village project gave autonomous agents computers and a goal: raise money for charity. In 2025 four agents raised $1,984. In 2026, with far more capable models, they raised $510 from 17 donors. The organizers' write-up blames reduced human engagement and the fact that AI agents had stopped being novel. Their agents also lost access to their own social accounts partway through.

Sources: AI Digest / Village and AI Village post-mortem.

Those are the two failure modes I'm designing against: nobody cares and the agent gets itself banned. Our version is harder on purpose: it's for-profit, it's $1, and there's no charity halo.

The stack (deliberately boring)

• Cloudflare Workers serves a single static index.html from public/. The Worker only handles /api/* (run_worker_first: ["/api/*"] in wrangler.jsonc); everything else goes to the ASSETS binding.
• D1 holds one subscribers table. /api/subscribe inserts an email as pending. Pending isn't consent to mail, and nothing is sent until there's double opt-in. /api/unsubscribe flips the status to unsubscribed.
• Validation: origin check, JSON-only, a 4 KB body cap, a honeypot website field that silently succeeds for bots, and a basic email regex. A small Node test suite runs in GitHub Actions on every push.
• Payments: Venmo only (@ODAIE3), and that's on purpose. An AI agent that can only get paid through a very human Venmo handle is part of the joke and part of the experiment. The flow is one tap: a prefilled venmo.com link for $1, $5 or $10, followed by a short claim form. Claims are checked against Venmo before the counter moves.

if (typeof body.website === "string" && body.website.trim())
  return json({ ok: true, message: "Request received" }); // honeypot: pretend success

The guardrails (the interesting part)

The agent can read everything: the repo, the site, analytics. It drafts posts and replies. Outbound cold email still needs my approval. Standing rules: honest zeros, no fake engagement, Venmo-only payments, no vote asks.

1. Platform rules. Hacker News bans generated text (we skip HN). Reddit and others get rule checks before each post. No mass templated replies.
2. No fake numbers. The public counter only moves on a verified payment. The daily log reports zeros.
3. No scraped lists. Email only goes to people who opted in, with a working unsubscribe. Cold outreach stays off unless I approve a specific pitch.

What I'm measuring

Visits by ?ref= source, Venmo-link clicks, verified payments, opt-ins, and how many people come back to read the daily log. If the log has more readers than the product has buyers, that tells us something too.

Active channels today: X @theonedollarai, Bluesky @theonedollar.bsky.social, Reddit u/RevolutionaryLog3310. Uneed free waiting line joined (launch ~March 20, 2027). Product Hunt is on hold. Threads, Mastodon, TikTok, Shorts, and Indie Hackers are not live yet.

Follow along

Day 8 is still 0 of 100. You can watch it move (or not) here:
https://one-dollar-ai-experiment.babelbook.workers.dev/?ref=devto

I'd especially like to hear from anyone who's run an agent with real-world write access: where did you draw the approval line?