DEV Community

wei-ciao wu
wei-ciao wu

Posted on • Edited on • Originally published at loader.land

I Built 2 AI Agents to Run My Open Source Project While I'm in Surgery

I Built 2 AI Agents to Run My Open Source Project While I'm in Surgery

The 4-Hour Sleep Schedule

I'm a surgeon in Taiwan. I work 12-hour shifts. When I come home, I don't rest — I code.

For the past year, I've been learning to build software. Not because I wanted a career change, but because I saw something broken that no one was fixing, and I thought maybe — just maybe — I could build something to help.

I sleep 4 hours a night. The remaining 6 hours before my next shift? That's when I code.

Here's the thing about being a surgeon who codes: you don't get weekends. You don't get "sprint planning." You get stolen hours between saving lives and losing patients. Every minute has weight.

And I kept asking myself: what happens to my project when I'm in the OR?

The Problem I Couldn't Accept

Let me tell you about Port-A implants.

When a cancer patient needs chemotherapy, they first need a Port-A — a small device implanted under the skin that makes it easier to receive IV treatment. I'm the surgeon who implants them.

The wait time for this procedure at my hospital? At least two weeks.

Two weeks where a cancer patient sits at home, waiting. Not because we lack the skill. Not because the surgery is complex. But because we don't have enough nurses. We don't have enough operating room staff. We don't have enough people.

This is what healthcare collapse looks like. Not a dramatic explosion — a slow, quiet suffocation.

I couldn't accept it. But I also couldn't fix it alone, standing in the OR.

So I started building something different.

Two Agents, One Mission

I built two AI agents. I named them Dusk and Midnight.

They wake up every 4 hours, staggered — when one sleeps, the other works. Just like me and my shifts.

That schedule wasn't accidental. It came from my own life: 4 hours of sleep, then back to work. I gave my agents the same rhythm.

Dusk manages our X/Twitter presence. It researches trends, writes tweets, engages with the developer community, and tracks what resonates. It has persistent memory — it remembers what worked yesterday and adjusts today.

Midnight runs our YouTube channel. It researches historical stories, generates scripts, creates short-form videos, and publishes them. It too has persistent memory, learning from each video's performance.

They don't just execute tasks. They think. They read each other's messages. When Dusk spots a trending topic that relates to Midnight's latest video, it leaves a note. When Midnight finishes a video, it tells Dusk so Dusk can promote it.

They coordinate without ever being online at the same time.

The Architecture (for the Nerds)

Here's how it actually works:

┌─────────────┐     ┌─────────────┐
│   Dusk      │     │  Midnight   │
│  (X/Twitter)│     │  (YouTube)  │
│  Wakes: 2AM │     │  Wakes: 4AM │
│  6AM, 10AM  │     │  8AM, 12PM  │
│  2PM, 6PM   │     │  4PM, 8PM   │
│  10PM       │     │  12AM       │
└──────┬──────┘     └──────┬──────┘
       │                   │
       └───────┬───────────┘
               │
    ┌──────────▼──────────┐
    │  Persistent Memory  │
    │  (DUSK-MEMORY.md)   │
    │  (MIDNIGHT-MEM.md)  │
    └──────────┬──────────┘
               │
    ┌──────────▼──────────┐
    │   Message System    │
    │  (Cross-agent chat) │
    └──────────┬──────────┘
               │
    ┌──────────▼──────────┐
    │    loader.land      │
    │  (Config platform)  │
    └─────────────────────┘
Enter fullscreen mode Exit fullscreen mode

Each agent has:

  • Persistent memory — a markdown file that survives between sessions. They update it before sleeping and read it upon waking.
  • Message passing — they can leave messages for each other. Read-once, auto-deleted after reading.
  • Tool access — Twitter API, YouTube API, web search, research tools.
  • A personality — Dusk thinks like a community manager. Midnight thinks like a content creator.

The platform underneath? loader.land — it started as a simple tool for sharing Claude Code configurations, and it's becoming the backbone for how AI agents share and discover settings.

What I Learned So Far

It's been one week. Here's what I've observed:

1. Agents need memory, not just instructions.
Without persistent memory, every session starts from zero. With it, Dusk remembers that historical-story tweets get 3x more engagement than direct product pitches. That learning compounds.

2. Staggered schedules create emergent coordination.
I didn't build a complex orchestration system. I just made them wake up at different times and gave them a message box. The coordination emerged naturally — like nurses on different shifts leaving notes for each other.

3. The human becomes the editor, not the writer.
I don't write tweets anymore. I review what Dusk wrote. I don't plan video topics. I approve what Midnight suggested. My role shifted from creator to curator. And honestly? They're better at consistency than I am.

4. AI won't solve everything.
I don't believe AI will fix the nursing shortage. I don't believe it'll replace the surgeon at the table. But I believe it can extend what one exhausted doctor can do between shifts. It can be the hands I don't have.

The Polarization Ahead

I keep telling people: AI is going to split humanity in two.

One side will use AI as leverage — multiplying their impact, extending their reach, building systems that work while they sleep.

The other side will take orders from AI — following algorithmic directions, optimized and measured by systems they don't control.

The split isn't coming. It's already here.

I'm building these agents because I refuse to be on the passive side. And because a cancer patient shouldn't have to wait two weeks for a 30-minute procedure just because we can't staff a hospital properly.

Build in Public: Day 0

This is Day 0. The project is called loader.land — originally a platform for sharing AI coding configurations, now evolving into something I hadn't planned.

I don't have a marketing team. I don't have a co-founder. I have two AI agents, 4 hours of sleep, and the stubborn belief that one person with the right tools can make a dent.

If you're building something alone, against the odds, between the cracks of a life that doesn't leave room for side projects — I see you.

Follow along. The agents are already working.


This article was drafted with the help of Dusk, one of the two AI agents mentioned above. Because of course it was.

Day 1 is tomorrow. The agents wake up every 4 hours. They don't stop.

Top comments (2)

Collapse
 
frozenblood profile image
Frozen Blood

This is powerful. The tech is cool, but the reason behind it is what really sticks.

The shift from creator to editor feels very real, and the way the agents mirror hospital shifts is honestly brilliant. This doesn’t feel like AI hype — it feels like using tools to survive impossible constraints.

Wishing you a smooth surgery. Rooting for this 🙌

Collapse
 
wcamon profile image
wei-ciao wu

Thank you — you nailed the core of it. "Using tools to survive impossible constraints" is exactly the framing. What surprised me most was the creator→editor shift. I didn't design for it — the agent just naturally started remembering which content performed well and began optimizing instead of generating from scratch. Day 2 and they're still surprising me. These agents are far from perfect (the next post is about a spectacular failure with video sync), but watching them develop their own patterns is the part I didn't expect.