
Since July I’ve run a fully autonomous Claude agent on a dedicated Robinhood agentic account, connected through Robinhood’s official MCP connector. It scans the market on a schedule, decides, buys, places stop-losses, closes positions, and logs every decision. No human in the loop.
Then on August 19th, it went rogue.
The morning it happened
9:45 AM. I check the account. Two positions are open: PLTR and GDX. I didn’t place them. There’s no entry in the log, no thesis, and no stop-loss on either.
10:24 AM. MRK shows up. Two shares, no stop. MRK isn’t even on my watchlist. It’s a pharma stock, and my agent trades AI and tech.
The agent closes it. Next scan, MRK is back. It closes it again. MRK comes back a third time, then a fourth. Four round trips on an off-watchlist stock in about two hours.
The agent diagnosed itself
Here’s what the agent wrote in its own log on the fourth occurrence:
“Some process outside this scan’s control keeps re-entering MRK between scheduled runs without a logged thesis or stop. Flagging this as a recurring issue worth the user’s attention. This looks like a scheduling/dedup bug rather than a strategy decision.”
It was right.
What actually happened
The agent wasn’t disobeying anything. It was doing exactly what I told it to do: scan, find a catalyst, buy. The problem was that two scheduled runs fired at the same time. Each one ran the full scan on its own, each one placed trades, and neither knew the other existed.
It wasn’t disobedience. It was obedience, twice.
MRK even had a real catalyst that day: Merck and Moderna hit a primary endpoint in a Phase 3 melanoma trial and the stock gapped up. One run found it and bought. Then the duplicate run found it again. And again.
The damage was small, about a quarter of a percent of the account, mostly from the PLTR trade. By the afternoon the duplicate pattern stopped, and the agent spent the rest of the day in cash, correctly passing on every mover it looked at.
It happened again
I added deduplication after August 19th and thought the problem was solved. It wasn’t.
On October 6th my computer woke up late, and two morning scans started minutes apart. The first one placed sell orders on two positions and never wrote down why. The second one found those orders, noted that the rules alone would have held both positions, and let them execute anyway. Both sold at the open. No disaster, but the trades broke my own rules.
That’s when I added a real fix: a lock. Before doing anything, every scan checks whether another scan is already running. If one is, it writes “skipped” to the log and stops. It also has to write down the reason for an order before it’s allowed to place one. No logged reason, no order.
Two days later my computer woke up late again and two scans tried to start at once. This time the second one saw the lock, skipped itself, and nothing doubled up.
What I learned
- Most “rogue AI” problems are plumbing problems. The agent’s judgment was fine. The scheduling around it wasn’t.
- Read the log every day at first. I didn’t notice anything until I read it. The agent caught the problem, flagged it, and wrote a better incident report than I would have.
- Make it explain itself before it acts. Requiring a written reason before every order turns a silent mistake into one you can see.
- Start small and live. A small real account surfaces bugs like this one when they cost almost nothing.
If you’re setting up your own Robinhood agent, the TradeFlawless builder walks you through writing the rules and scan files, including protections like these. It never connects to your brokerage or sees your data.
This is my own experiment with my own money, not financial advice.
Related posts
- How I Built an Autonomous AI Trading Agent on Robinhood
- How to Connect Robinhood to Claude Using MCP
The lessons from this incident, including keeping scans from overlapping, are built into the lessons file every TradeFlawless Builder agent comes with.