It runs a loop — screen the market, analyse a candidate, size it, route the order, monitor the position, exit — and the only interesting engineering problem is what it must refuse to do. You define capital, markets, risk limits, and how much autonomy it gets: research-only, approve-every-order, or bounded automation. It cannot exceed the grant.
It starts on a paper account. Live trading only happens after you explicitly connect a broker.
Things I got wrong that might be useful to anyone building agents that touch real systems:
Our stop control rendered only while a session was "running". A session can be alive but unhealthy — expired broker auth, errored runner, halted risk — and in those states we withheld the stop button. So the control existed in every situation except an emergency. One paper session sat unstoppable for 218 hours while the UI advised stopping it.
A reset button that rendered inside a scroll container, so on desktop the wheel scrolled the page and it could never be reached. Rendered and unreachable is identical to missing.
A disconnect that refused when there was nothing to disconnect — aborting before cleanup, which made the dead record permanent.
Sizing off a failed price read. If a quote call returns the last good tick instead of an error, the agent sizes against a number that doesn't exist. Fail closed on reads, always.
The general shape: guards get written from the happy path, and the broken states are by definition the ones nobody pictured. In most software that costs an afternoon; here it holds a position.
What I deliberately don't have: a verified live performance record. I publish the evidence standard I'd have to meet before claiming one (quantsignals.xyz/fst/proof) and I haven't met it. If someone shows you a win rate for an agent like this, ask what the denominator was.
Happy to answer anything about the architecture, the broker integrations (8 of them, each one its own dialect of pain), or the risk model.