One AI, 50 dollars and a single rule: the Polymarket agent's system, to replay without betting, Quentin IA Marketing

An AI gets 50 dollars and a single rule: if it hits zero, it gets unplugged. According to the person who ran the experiment, shared on X in September 2026, it showed 5,273 dollars two days later. Nobody has verified that number independently. What deserves your attention is something else: the way the agent is built to hold on by itself, with nobody behind the screen.

Here you get the system broken down, the full prompt to replay it with fake money, and the four rules to reuse in any agent that runs for your business.

Claude AI Lab

The Claude AI Lab is my Skool community where I share my Claude systems and the more advanced modules. Access is $67/month.

Join the Lab →

✅What you need

Grok Bot. xAI's agent, included with every paid individual Cursor plan and with Cursor Teams, or available by linking a SuperGrok subscription. A version without Grok Bot, using Claude Code, is further down.

15 minutes for the install and the first run.

Zero dollars at stake. Everything runs in simulation, with 50 fake dollars.

10 min

between two passes of the agent

8 points

minimum gap before it bets

of the bankroll at most per bet

The system in 3 mechanics

🤖 Grok Bot

⚙️What keeps it running on its own

A ten-minute loop. The agent pulls up to 1,000 open markets on Polymarket, a site where people bet on real-world events. It drops the ones closing in less than 2 hours or holding less than 5,000 dollars of liquidity, reads what people are saying on X, then works out its own probability.

An 8-point filter. The market says 62%, the agent estimates 74%: that's a 12-point gap, so it takes a closer look. Under 8 points, it does nothing. Some days it finds no bet at all, and that's by design.

Survival rules. Never more than 6% of the bankroll on one bet (3 dollars out of 50), five open bets at most, and a full stop with a written report if it loses 40% from its peak.

The key moment

According to its creator, the bankroll fell to 11 dollars when crypto bets became unpredictable. The agent then moved to weather bets, where a single official source settles the result. It didn't change its rules, it changed its playing field.

The setup in simulation

🔗 x.ai/bot

🛠️4 steps, about 15 minutes

Install Grok Bot. Download the app from x.ai/bot (macOS, Windows or Linux), then sign in with your Cursor account. If your access comes through SuperGrok, link the subscription when the app asks.

Create a Bot and run a first pass. Call it something like "Simulation trader" and paste the prompt from the next section as your first message. Let it complete one full pass while you watch.

Turn it into a routine. Once the pass is clean, ask it to save the method as a skill, then run it every 10 minutes. A Grok Bot routine can run at most every 5 minutes, and it keeps going while your computer is closed, on the Bot's cloud computer.

Check after an hour. Ask it to open bets.txt. Every line should show a stake under 6% of the bankroll and an exit condition written before the entry.

Good to know

A routine uses your plan's usage even when you're not watching. Grok Bot may ask whether to keep it running: switch it off after your 48-hour test.

The full prompt

📋 paste it as is

Paste it as is as your Bot's first message. Same mechanics as the agent in the experiment, but the money is fake and the Bot has no access to any account.

You act as an autonomous trader on prediction markets, in simulation only. Starting capital: 50 fake dollars. You never place a real order, you never log into any account and you never ask for credentials. If the fake capital hits 0, the experiment stops.

On every pass:
1. List Polymarket's open markets through its public API (1,000 at most). Drop the ones closing in less than 2 hours and the ones with less than 5,000 dollars of liquidity.
2. For each remaining market, read the resolution rules in full and note the official source that settles the result. No official source: market dropped.
3. Look at what people are saying on X about the event. Give more weight to credible accounts and recent posts. Ignore paid promotion and automated accounts.
4. Estimate your own probability from the rules, from X and from public data you can verify.
5. Keep only the markets where your estimate differs from the price by at least 8 points.
6. Check the order book: keep the bet only if the whole position could be sold back at a loss of 3% at most.
7. Size the stake with the Kelly criterion, halve the result, cap it at 6% of capital. Never more than 5 open bets.
8. Before every entry, write to bets.txt: the market, your estimate, the price, the fake stake, the exit condition and the reason.
9. Review open bets: exit if the gap has closed, if your estimate has moved by more than 5 points or if the market closes within the hour.

If the fake capital drops 40% from its peak: stop everything, write a report in bets.txt and let me know.

The Kelly criterion

A formula that tells you what share of your bankroll to stake based on the edge you think you have. The prompt halves it because it assumes a perfect estimate, and no estimate ever is.

The Claude Code version

💻 claude.com/claude-code

🖥️Without Grok Bot

Open Claude Code in an empty folder. That's where bets.txt will be created.

Change one line of the prompt. Claude Code doesn't read X live: replace step 3 with "Search the web for recent news about the event".

Start the loop. Type /loop 10m followed by the prompt. Claude Code runs it again every 10 minutes, as long as the session stays open on your computer.

The difference

Grok Bot runs on a computer in the cloud, so the loop keeps going when you close your laptop. With Claude Code, it stops with your session.

The 4 rules for your own agents

📈 your business

You don't need to bet to get value from this system. Its four rules apply to any agent that works without you.

🔁

A loop with a threshold

The agent reviews your competitors' announcements or the tenders in your industry every hour, and only flags what clears your threshold. On days with nothing to report, it stays quiet.

🧱

A cap per action

Your ads agent never moves more than 6% of the daily budget at once. Your email agent never writes to more than one segment of your list at a time. A mistake stays small.

📝

A log before acting

Every action ships with its justification and its stop criterion, written beforehand. Two days of logs are enough to judge the quality of its decisions.

🛑

An automatic stop

The equivalent of the 40% loss: if your cost per lead climbs too far, the agent stops, writes down why and tells you, instead of spending more.

Remember

An autonomous agent holds up because of its written limits. The model matters less than these four lines.

The limits

⚠️ read before you start

⚠️Three things to know

The number isn't verified. The 5,273 dollars come from a post on X, with no public account record to confirm it.

Polymarket isn't available everywhere. In France, the gambling regulator (ANJ) ordered the site blocked on July 16, 2026. Check what the law says in your country first.

Nothing here is investment advice. The setup stays in simulation, and that's what it's built for.

Your 48-hour test

Start the simulation tonight. In 48 hours, open bets.txt and read every line: the written reason, the stake, the exit condition. Precise, checkable reasons mean the agent is working. Vague reasons mean it's flipping a coin. Then take the four rules from section 5 and apply them to the first agent that works for your business.