I built a system of AI agents that runs my money, my content, and my attention — on a schedule, not on demand. Here's the architecture, piece by piece, and what it actually looks like running in production.
Mine runs on a VPS — a small virtual server in a data center. Always powered. Always online. Nothing fancy; the same kind of machine people use to host a blog.
This is the environment where everything else happens. Think of it as the body the agent inhabits.
Not a chatbot. Not always listening (but always available). A scheduled program that starts when it's supposed to, reads its context, makes decisions using an AI model, and writes results to files.
The whole runtime is one command:
A markdown file called CLAUDE.md. The agent reads it at the start of every session — before acting on anything. It contains identity, values, rules, ways of thinking.
Mine is 731 lines of my own investment reasoning, written down. Each line a decision I made once and want the agent to reproduce.
Agents don't remember yesterday by default. So every run writes its work to shared markdown files — portfolio state, open theses, decisions, the last briefing.
Tomorrow's wake-up reads those same files and knows what the world looked like the last time it ran. Memory, simulated.
Five scripts run at 4:30 AM every weekday. They pull from specific sources — insider trading disclosures, SEC filings, macro calendar, institutional 13F filings, congressional trade reports. Each writes a markdown file the agent reads during its morning wake-up.
The agent doesn't browse the web in real time. It reads a pre-digested snapshot. Fast. Predictable. Cheap.
When the agent has something worth saying — a briefing, an alert, a question — it sends a Telegram message. I read it on my phone like any other text.
No dashboard to check. No app to open. The agent comes to me when it matters; otherwise, it's silent.
Cron is a standard feature of every Linux system — it runs programs on a schedule. My agent has 31 scheduled jobs across a week: a morning briefing at 6 AM, a market open alert at 6:35 AM, a close review at 1 PM, weekly deep dives on Thursday afternoons.
Predictable. Auditable. Cheap. If something misfires, I know exactly when and why.
Every tick is a scheduled job firing. Tap any dot to see what happened.
Deterministic shell scripts — not AI. Each pulls from a source and writes a markdown file the agents will read later.
Every Sunday the financial agent screens the entire US market — a year of price history for every listed stock — and forces every survivor through one gate: name the edge, or pass. Every recommendation it makes is logged against the exact price where it would be proven wrong, and it must reread its own track record before making a new call.
See how it hunts →At the start of every session, the agent reads these files in order. They define who the agent is, what it knows, and what it's allowed to do. Change the files, change the agent.
Values, judgment, ways of thinking. The first file read at every wake-up. If the agent had a soul, this would be it.
Live positions. Rewritten at every close review so tomorrow's agent knows where it stands.
Thesis for every position. Entry rationale, kill-conditions, what would invalidate it. Read every morning so the agent can grade theses A through F.
Pulled fresh at 4:34 AM daily. Cross-referenced against the watchlist for signal. Refreshed from House Clerk + Senate disclosures.
Quarterly holdings of tracked managers. New positions, increased stakes, full exits. Signals with fiduciary-grade conviction.
Append-only log of every decision. When the agent explored X and chose Y over Z, it's here. Prevents the agent from re-litigating settled questions.
My agent is personal. It's built for the way I think, trade, and write. But the architecture — a file for identity, cron for rhythm, a shared filesystem for memory, hard rules for safety — works for almost any job you want an agent to do.
If you want one of your own, tell me what you'd want it to do. I'll reply honestly whether it's a good fit.