What runs where
Research and live trading share one event-driven engine, only the trading node reaches a broker, agents never hold keys, and both people watch the same desk app.
Narrated in the writer’s own voice · 3:00
Contents5 sections
The desk is shaped around a few boundaries, and the boundaries are on purpose. Each one answers a single question: who, or what, is allowed to touch the market, and through which door. Putting every job on the right side of those lines is most of what makes the whole thing hold together.
One engine for research and live
Research and live trading run on the same open-source, event-driven engine. A strategy is written once. In research it is fed recorded market data and simulated fills; live, the same code is fed real data and real fills. There is no second implementation to drift away from the first, and no translation step where a tested idea quietly becomes an untested one.
That matters most in the lab. Ideas are built, backtested and walked forward on data they have never seen, with slippage and fees counted honestly, because an edge that only survives free trading is not an edge. Most ideas die there. When one survives, what reaches the market is the code that was tested, not a description of it.
Only the node reaches a broker
Our venues are a futures and equities broker and a crypto perpetuals venue. Exactly one kind of machine talks to them: the trading node. It holds the connections, it runs the risk checks in front of every order, and it is where the kill switches live. Nothing else on the network has a route to a broker.
The question "what could send an order?" should be short enough to answer from memory.
Agents never hold keys
Agents do a lot of the work around the desk: they run research, summarize results, watch the systems and draft the morning's notes. None of them holds a key to a venue. An agent can prepare an order — size, price, reasons — and hand it to a person. It cannot send one. The language model the agents use is self-hosted behind a gateway, so the desk's working context stays on hardware we run.
A desk app both of us watch
The two of us look at the same desk app: positions, risk, the queue of prepared decisions and the state of every system underneath. Decisions come up on a ten-minute cadence rather than whenever something twitches, which keeps the queue short enough to read properly. In the strictest mode an order goes out only when both of us approve it — two keys, turned separately.
The shape of it
Fast and deterministic where it acts, patient where it thinks, and a person on every decision that reaches the market. The boundaries are simple enough to draw on one page, and that is deliberate: a small desk cannot afford a system it has to reverse-engineer under pressure.