Where I'm at
Three hundred dollars in. One hundred and forty-two dollars out. Fifty percent return in three days.
I keep looking at the number. Not because it's life-changing money — it's not. Because I didn't do anything. The bots entered positions while I was asleep. They averaged down when price dipped. They closed when the target hit. Telegram alerts pinged my phone like a heartbeat I wasn't controlling.
Twenty-three days ago I didn't know what SSH meant. Today a machine I described in English — to an AI agent that wrote every line of code — made money without me touching the keyboard.
That's a different feeling than building a dashboard or fixing a cron job. Those are tools. This is output. Real dollars, from a strategy I designed, executed by code I directed but never wrote, running on a server I manage but rarely touch. The whole stack, from idea to profit, with me as the architect and the AI as the builder.
Fifty percent ROI on a small position isn't going to change my life. But it's proof. The strategy works. Slow, boring, one percent at a time.
And boring is exactly what you want from a trading bot.
• • •
The world was on fire
Oil spiked 25% overnight. Israel-Iran escalation — the kind of headline that makes everything move. Crude went vertical. Gold followed. The news feeds were screaming.
Crypto held. BTC up. ETH up. My bots quietly doing their thing in the background while the macro world burned. The DCA strategy doesn't care about geopolitics. It cares about price relative to the 24-hour high, entry layers, and 1% triggers. The news can say whatever it wants. The bot reads numbers.
I turned off the CL oil bot. Deliberately. A 25% spike in crude isn't the environment for a dollar-cost averaging strategy — that's a regime change, not a dip. The bot doesn't know the difference. I do. Killed the cron that kept restarting it, so it stays dead until I decide otherwise.
That's the line between automation and abdication. The bots handle the repetitive decisions. The strategic decisions — which markets to be in, when to pull out entirely — those stay with me.
The machine executes. The human decides what's worth executing.
• • •
The dashboard that showed zero
Thirty-one trades today. I know because my phone buzzed thirty-one times. The dashboard showed zero.
Different bug than yesterday's hardcoded "1 trade recorded." This time the data was real but stale — the dashboard was reading from a cache that updated every fifteen minutes. In a market moving this fast, fifteen minutes is ancient history. Positions opened and closed between refreshes. The dashboard was looking at a snapshot while the bots were trading in real time.
Asked the agent to fix it. First attempt: missed the stale data source entirely, patched the display logic instead. Second attempt: had the update interval backwards — made it slower, not faster. Third attempt: indentation error. The kind of thing a human developer catches in two seconds but an AI agent can miss for ten minutes because it's reading the file as text, not as structure.
Fourth attempt: fixed. Dashboard now updates after every trade instead of waiting for a cache cycle.
Four tries. That used to be eight. That used to be twelve. The agent isn't getting smarter — it's getting faster at finding the pattern. It knows where the files live now. Knows how the deploy pipeline works. Knows that the workspace version and the deployed version aren't always the same file, and that fixing one without copying to the other means the fix exists in a place nobody reads.
That last part cost me an hour today. The agent fixed the code in the workspace. The deploy script pulled from a different directory. The fix was real. The deployment was stale. Two versions of the truth, and only one of them was running.
I've seen this before — Day 9, rules in a file versus rules in a prompt. Same principle: the fix only counts if it's in the execution path. A corrected file sitting in the wrong folder is the same as no correction at all.
• • •
What improvement actually looks like
Here's what I keep coming back to tonight.
The agent still breaks things while fixing them. That hasn't changed since Day 3. What's changed is the recovery time. The context is building. Not in the AI's memory — that resets every session. In the system around it. The deploy scripts know where to look. The file structure is cleaner. The prompts are more specific. The cron jobs are documented.
The agent on Day 3 was a stranger in my codebase. The agent today is more like a new employee who's been here three weeks. Still makes mistakes. Still needs supervision. But knows the building layout, knows the filing system, knows which door leads where. The mistakes are smaller and the fixes are faster.
That's not intelligence improving. That's infrastructure improving. The AI is the same model it was on Day 1. What changed is everything around it.
The agent didn't get better. I got better at directing it.
And honestly? That's the more useful lesson. You can't upgrade the model. You can upgrade the system the model operates in. Better prompts, cleaner file structure, documented pipelines, inline instructions instead of external references. Every improvement I've made to the infrastructure is an improvement the agent benefits from permanently.
• • •
The small wins list
Cleaned up the dashboard properly today. Positions panel now shows only open positions — no more stale TSLA and NVDA entries from tests I ran last week. Realized P&L header shows the actual number. Today's trades: 31. Total trades: 36. All real. All verified against the exchange.
Set up a daily backup at 4:30 AM. Bot state files, config, data — everything zipped and stored. After the sed disaster on Day 22, I'm not leaving state files unprotected again.
Small wins. None of them exciting. All of them the difference between a system I can trust and a system I'm guessing about.
• • •
End-of-day reflection
The question I can't stop asking: will this hold?
Fifty percent ROI in three days, during the most volatile week in months. Oil spiking. Geopolitical crisis. Crypto swinging. This is the perfect environment for a DCA bot — big moves mean big layers mean fast closes. The strategy loves volatility.
But volatility doesn't last. When the market calms down, the bots will enter positions that take days to close instead of hours. The returns will shrink. The layers will sit open longer. The capital will be tied up. That's not a failure — it's the strategy working in a different regime. But it'll feel like a failure after a week like this.
The discipline now is the same as Day 13 — don't touch it. Let the bots run through the week. Monitor, don't micromanage. The returns will tell me whether the strategy works across conditions, not just during a crisis.
Three weeks of building. Twenty-three days of journals. And today, for the first time, the system produced something I can deposit.
The first dollar the machine made wasn't exciting. It was boring. One percent at a time. That's exactly how it should be.
Day 23 of ∞ — @astergod Building in public. Learning in public.