GPT-6 Astra runs six trading agents: $700 to $6,310 in 11 nights
The author blinds each of six Astra agents to one key input (price, position size, win/loss), lets them communicate only through verdict cards, and over 11 nights filters 4,561 setups down to 41 trades.
OPENAI'S GPT-6 ASTRA RUNS SIX TRADERS ON MY $700 PAPER ACCOUNT AND FIVE OF THEM ARE NOT ALLOWED TO SEE THE PRICE. IT IS AT $6,310 AFTER 11 NIGHTS everybody building agents gives all of them the same screen and wonders why they agree with each other. i built mine the opposite way, where every agent is blind to the one thing that would make it stupid: -> PROBE only gets volume and order flow with the price column deleted, so a pretty candle can never talk it into anything -> PRISM sees who paid for the move but not how big the move was, because size is what makes people chase -> TALLY does the sizing and is never told whether the last trade won or lost, so revenge sizing is impossible -> HATCH writes the exit price before the entry exists and it is the only agent allowed to close a position -> LEDGER writes one line for every dead setup, and the other five have to read the last 20 lines before they are allowed to speak -> WARDEN sees no chart at all, only the five verdicts, and answers one question, has this exact shape killed us before warden runs on a prompt that is 940 words long while the other five are about 120 each. it vetoes 9 out of every 10 setups and it has never once been shown the price of anything. in 11 nights the desk processed 4,602 setups at 0.4 seconds a round, killed 4,561 of them and let 41 trades reach the account. the $700 paper account reads $6,310 this morning. the agents never talk to each other, they pass a verdict card and nothing else, and that single rule fixed more than any model upgrade did. the six prompts and the verdict card format are in the article below, start with the warden.
This post quotes another post: x.com/starmexxx/status/2098293921994465300