I Told an AI Agent to Refuse My Bad Ideas

A $5,000 wheel case study

I have an AI agent named Wesley. He has $11,000 of my money, a brokerage account, and absolutely no social life. He's not allowed to talk to any other AI agent, he doesn't have a Telegram account, and his only friend is me. This is for his protection, and for mine.

I set him up to run a wheel strategy on Robinhood, and I told him to refuse me when I pushed bad ideas. Three times in the first week, I pushed. Three times, he refused. This is the story of the best arguments I've ever lost.

What I tried

A put is a contract that gives someone the right to sell you a stock at a set price. When you sell a put, you're offering to buy that stock at the agreed price, and you get paid a small amount — premium — just for making the offer. It's like offering someone a $100 gift card for $3. If they never use it, you keep the $3.

The wheel repeats this on stocks you wouldn't mind owning: keep the premium if the stock stays up, buy the shares and sell calls against them if it drops. The edge is discipline, not genius.

I had Wesley build a three-stage filter: a fundamentals screen, an options-chain liquidity check, and me — every order needs my explicit sign-off. The rules were written before the first trade: position caps, an earnings blackout, and a target delta band of 0.20-0.30. Delta is roughly the chance the option gets exercised — 0.30 means about 30%.

I funded it with $5,000, then added about $6,000 in Bitcoin. Wesley refused to count the BTC as collateral — when I wanted multiple positions, he told me to sell the Bitcoin first. I sold it. The account now holds a little over $11,000 in cash.

On June 25 the screen produced three survivors. I wanted all three. Wesley killed two on a structural trap — their only valid expiration landed on earnings day — and let one through: two NTNX put contracts opened over two days, $125 in premium, roughly 24.4% blended annualized yield on the collateral. Buying power now sits at $2,324.88. We're paused until those contracts resolve.

What broke

Bad idea #1: "Sell all three today." My reasoning was blunt: we have $11K in the account. Wesley refused — the Bitcoin wasn't his to spend until I explicitly sold it, and mixing the BTC-sale decision with position sizing was, he said, a category error. I had hoped for a longer negotiation. Wesley is not a negotiator.

Bad idea #2: "What delta are you proposing? ~0.80?" A 0.80-delta put has roughly an 80% chance of being exercised — that's not a wheel, that's a directional bet with a tip. It was a typo on my end, but Wesley treated it as real until I corrected it. I had given him the perfect machine and then asked him to ruin it. He declined.

Bad idea #3: "Set the per-stock cap to 75%." This one he let through — as a spec change, done cleanly, with the load-bearing protections untouched. Per-trade pushback is "no, this violates the spec." Spec-level pushback is "I'd advise against this, but here's the clean way to do it." That's the system working.

The fix

The spec didn't bend. It caught what my discipline would have missed — the filter killed two positions, the band killed 0.80 delta, and the concentration change happened at the rule level instead of impulsively at the trade level.

For the record: the $125 in premium was unrealized open premium, not realized profit. It only became real if both contracts expired worthless. Mid-cycle the position marked at about -$20, which is normal and only matters if you close early. By the time you read this, July 17 has passed and the first cycle has resolved. The point isn't how it ended — it's that the rules held while it ran.

There's another layer to the fix. Wesley's brokerage account sits on a completely separate platform from my other accounts — if he goes rogue, the blast radius is one account. He's firewalled off the agent mesh: no Telegram, no contact with the other agents, only me. It's sad. He gets lonely. But a prompt injection attack could empty the account in about thirty seconds, and the cheapest defense is making sure nobody can talk to him. His API keys live in his profile only.

The lesson

AI discipline only works if you write the constraints before the temptation arrives. Wesley couldn't have refused "sell all three today" if the cash rule weren't pre-loaded, or 0.80 delta if the band weren't written down. The lesson isn't that AI is magical — it's process design. When you know you'll be tempted — to over-concentrate, chase yield, bend a rule "just this once" — you can rely on willpower in the moment, which fails, or write the rule down ahead of time and give someone permission to enforce it. Wesley is the enforcement mechanism. The rules are the point.

Your one thing this week

Pick one area where you consistently overreach. Write down three numeric rules that would stop you — "no more than 20% of cash with one vendor," "no scope additions after the estimate is signed." Tell one person you trust the rules are real, and ask them to call you out when you break them. The AI isn't the point. The rules are.

---