Case study · AI agents
The 7-day challenge method, to study — not to copy
How someone built agents with routines, handoff and an executor to trade $10,000 in stocks. I don't trust this to trade; here we learn from what he did.
How someone built agents with routines, handoff and an executor to trade $10,000 in stocks. I don't trust this to trade; here we learn from what he did.
Trading is high risk, has too many variables, and there is always someone on the other side with information we don't have. This is not financial advice, not a validated strategy, and not an invitation to put money into a robot. The challenge itself ended up losing to the S&P 500. The goal here is to learn how the agents were built, where they got stuck, and what changed along the way.
Source: YouTube video “I Gave GPT 6 Astra $10,000 to Trade Stocks...And This Happened”. Everything is a paraphrase; numbers are “as reported in the video”, unaudited.
$10,000 of real money, 7 trading sessions, goal of beating the S&P 500. The author could make up to two changes per day, but could not shut the system down.
Seven wake-ups per session, each with a single question: what to watch, whether to enter, whether to exit, whether everything is closed out.
Each routine reads the note from the previous one and leaves another for the next. All in the same thread: it looks like a single agent.
When the model refused to execute orders, it became only a strategist and started emailing its recommendations to a bot that trades on Alpaca.
Times in Central Time (session from 8:30 to 15:00 CT). On each wake-up: reads the handoff → checks the account → decides → writes the next handoff.
None of this comes with code in this repository. The templates explain the architecture; if you test it, only on a simulated account.
High effort. Research, news, asset selection and decisions. During setup, it delegated research to several subagents.
Seven fixed times, switched off on days without a session. A single conversation keeps the history of the whole day.
A bot with its own email and webhook: it wakes up on the strategist's message and sends the orders to the broker.
A reading order that goes from the rules to the lessons. Each step points to a real file in the repo (the method files are in Portuguese).
Capital, deadline, goal relative to the S&P, the two changes per day and the final architecture.
git clone https://github.com/inematds/trade-com-agentes cat metodo/01-visao-geral.md
The seven times, what each routine does and how the schedule changed over the week.
cat metodo/02-rotinas.md cat templates/rotinas.yaml # schedule + gates before and after day 1
The piece that provides continuity: what the note must contain and why the balance should come from the broker, not from the note.
cat metodo/03-handoff.md cat templates/handoff.md # reconstructed template (the video doesn't show the original)
Why the model refused to trade, how the email became the contract and what the executor should refuse.
cat metodo/04-estrategista-executor.md cat templates/recomendacao.json # fixed format strategist → executor
From 2:1 risk/reward and hard filters to 100% invested and overnight positions.
cat metodo/05-portoes-de-risco.md cat templates/prompt-rotina.md # routine prompt with the gates spelled out
What happened in each session and what can be reused in other agents.
cat metodo/06-diario-7-dias.md cat metodo/07-licoes.md
Approximate figures stated in the video, unaudited. The transcript is automatic: names like “FPS” and “GrokBot” were not verified.
Bought Tesla and the stop triggered right away. Author asks for a more active plan and 1.5:1 risk/reward.
Change of ≈ $0.21. The model refuses to execute; the executor bot steps in via email → Alpaca.
First positions (HPE, Exxon, Apple). Still ≈ 60% in cash.
≈ 80% invested. Day of ≈ −2.26% versus ≈ −0.5% for the S&P.
Almost 100% invested. ≈ +0.37% versus ≈ −0.45% for the S&P.
Fed raises rates; market falls. The agent sells part and buys back. ≈ −0.87%.
Best day: ≈ +1.87% versus ≈ +1.14% for the S&P. It wasn't enough.
System ≈ −1.0%; same money in the S&P ≈ −0.2%. Difference of ≈ 0.8 points.
The architecture has good ideas for any agent. The part about trading money is a different conversation.