Investor desk
Gordon's desk: a 14-agent research team with a risk gate that can veto my own ideas
Gordon used to write one letter on a Sunday. Now he runs a desk: thirteen specialist agents covering news, earnings, macro, charts, value and risk, a red team paid to disagree, thesis cards with kill criteria, and a risk gate that blocks any proposal that breaks the rules I set. It never trades. It argues.
The first Gordon was one agent with a strong personality and a weekly letter. Good for arguing with over coffee. Not good enough to trust with a decision. So I built him a desk.
The team
Gordon is the portfolio manager. Thirteen specialists report to him, each with one job and one output file:
- Wire reads the news. Ledger reads earnings. Macro keeps the regime view.
- Trend, Momentum, Candles and Volume each read the charts their own way, and their votes are combined into one consensus.
- Alpha hunts for setups. Movers watches what moved and why. Analyst digests what the professionals are saying.
- Value keeps a watchlist of things that look cheap, with a margin of safety and a reason why the market might be wrong.
- Bear is the red team. His only job is to explain why the rest of the desk is wrong.
- Risk measures the book against the limits.
The rules have teeth
Most AI investing projects stop at “the model said buy”. This one is built around what stops a bad idea.
A mandate. One written document with the agreed shape of the portfolio and the limits. Gordon never invents a new allocation. He executes, monitors and proposes amendments, each ending with “Your call.”
A risk gate. Every proposal runs through a check against the limits before it reaches me: position size, concentration, volatility, drawdown. If it fails, it is vetoed, and the desk has to come back with a smaller or different version. My own ideas go through the same gate.
Thesis cards. Every holding has a card: why we own it, what has to happen, what would prove it wrong, and a date by which it should have worked. A card past its date gets flagged, not forgotten.
Validated rules. If a chart pattern is going to be used, it is tested across the whole universe and across time first. Most do not survive. That is the point.
It grades itself
Every call the desk makes is logged with the date and the confidence behind it. Each week the system scores which desks were right, checks whether “80% confident” actually comes true about 80% of the time, and breaks the results down by what drove them. Desks that are systematically wrong get a written lesson they read before their next run.
Where it lands
A short advisor note every weekday and a full deep dive every Sunday, both on a private dashboard and in Telegram. Positions and prices are read from the brokerage through a read-only connection and checked against a second source. Nothing places, changes or cancels an order. Ever.
What broke
- One confident agent is not a process. The original letter had conviction but no way to check itself. The fix was not a softer prompt. It was the red team, the gate, the thesis cards and the scorecard.
- Free data is only nearly right. The desk now checks every closing price against the brokerage before it grades a single call, and flags any gap in its notes.
What it is not
A personal research tool. Not financial advice, not a recommendation to buy or sell anything. It makes my own thinking more structured, and it is very good at telling me no.