Multi-agent trading system

The question

I wanted to know what it takes to let an AI agent act with real money and still trust the result. Paper trading doesn't answer that, because nothing is at stake. So I set up a small real account and built the controls around it.

This is an experiment in governance and engineering. It isn't investment advice, and it isn't a fund that takes anyone else's money. The public site is Collins Capital.

How it's organized

The system has three roles, modeled on how a real fund separates them.

• The fund manager is an AI agent. It researches and places trades inside a written risk rulebook. It's the only part of the system that can trade.
• The board is six language models that review the manager's reports and advise the owner. It has no brokerage access.
• The owner is me. I hold the only authority to change the rulebook or step in.

The trading tooling

The fund manager works through Python and SQLite tooling built for safe execution:

• Idempotent order handling. A retried order can't be placed twice, and every order is read back from the broker to confirm what happened.
• Stop-based position sizing. The size of a position follows from where its stop is.
• A performance journal. Every action is recorded, so the board and I review facts and not the agent's summary of them.

The board meeting

1. GatherReport, rulebook and journal, read-only
2. RedactIDs replaced, then a leak check
3. Round 1Six seats review independently
4. Round 2Each responds to the others and votes
5. Chair memo15 sections and an action register

Each seat is a different model from a different lab, called through OpenRouter, so the board doesn't share one model's blind spots.

The vote answers one question: does the owner need to step in? There are three possible answers, from "no owner action" to "urgent owner review." Poor performance alone can't trigger the urgent vote. Only a rule breach or a control failure can.

Every memo ends with an action register, and the next meeting starts by reviewing it. A full meeting costs about $1 or less in model usage.

Rules that keep it honest

• Advisory only. The board can question, analyze and escalate. It can't trade, veto, set limits or rewrite the rulebook.
• No hindsight. Decisions are judged on what was known at the time. A later price move doesn't prove a decision right or wrong.
• Small samples stay small. The board has to say whether it's looking at one observation or a pattern, and can't label skill from a handful of trades.
• A failed vote isn't invented. If a model's reply is empty or cut off, it's retried once and then recorded as "no valid vote."

Privacy

Before anything leaves my machine, account numbers and order IDs are replaced with stable labels such as [order-1], so a model can refer to the same order in both rounds without seeing the real ID. A second check scans the outgoing text and stops the send if anything remains. Each request also tells the provider not to retain the data.

The desk display

A companion project, the desk display, shows the account on a 7-inch screen driven by a Raspberry Pi Zero 2 W. It reads the account through a read-only connection and can't trade.

• A privacy view shows percentages only, for when other people are nearby.
• When a number isn't available, the screen says so and never substitutes sample data.
• Charts are hand-written SVG with no frontend packages, to fit in 512 MB of memory.

Testing

The board has 25 automated tests that run on fake files and a fake model client, so they need no network, no API key and no real account data. The desk display has 60 tests. Its suite covers the privacy view, missing values and market hours.

Still open

The desk display hasn't been through a 24-hour run on the physical Pi yet, and it still needs a reliable source for historical performance.