How the trading swarm justifies a trade.
The fastest way to build an agent that loses money is to let it trade on a signal. A signal is an opinion. An order is a commitment. The interesting engineering in the trading desk lives in the gap between the two — and in making sure that, after the fact, you can always answer the question "why did we own this?"
Paper account. Everything below runs on a simulated (paper) trading book. It is an engineering demonstration of an agentic system, not investment advice and not a live-money track record.
Signals are ranked opinions, not instructions
The research and signal bots produce a ranked list of candidates. Each candidate carries more than a ticker and a direction — it carries the reasoning that produced it: the market data and the alternative-data signals from World Intelligence that combined to surface it. That reasoning is not decoration. It is what makes the eventual trade auditable.
Crucially, the signal layer cannot place an order. It can only nominate. This is the same separation-of-powers idea that runs through the whole platform: the layer that has an opinion is not the layer that acts on it.
Risk sits between the signal and the order
Every nominated candidate passes through sizing and risk gates before it can become an order. Position sizing turns a ranked opinion into a specific share count given the account's constraints. Risk checks can veto outright — concentration limits, exposure caps, and sanity bounds all live here.
This is the step most "AI trader" demos skip, and it is the step that matters. Without it, the system chases every idea at full size. With it, the desk respects limits even when a signal looks irresistible.
Execution is a connector, and it is honest about which book it is
Orders route through a brokerage connector — the same connector-runtime pattern every other application on the platform uses. There are two ledgers, kept strictly separate: a paper book and a live book. The paper book proves the loop end to end. The live book is gated behind explicit human sign-off; no signal can promote itself from simulated to real.
Every fill is written to a ledger with a link back to the signal that triggered it. That backlink is the whole point. Months later you can pull any position and trace it to the reasoning that put it on — no black box, no "the model decided."
The day closes itself out
After the close, a post-event workflow takes over. It pulls the session's figures — profit and loss, leaders, the full tape of buys and sells — assembles a dashboard and a deck, narrates it, and renders a recap video. No human edits the clip. It is the platform's video workflow pointed at real results.
Why build it this way
The architecture is deliberately boring where it counts. Opinions are separated from actions. Risk is a gate, not a suggestion. Paper and live are different ledgers, not a flag. And the system is accountable by construction: every trade points back to its reason, and the day documents itself.
That is the difference between an agent that trades and a trading desk you could actually let near money — once it has earned it on paper.