Paper Lantern
AI Systems Engineer

I work with AI agents every day.
Here's what I've learned so far.

Multi-agent orchestration, AI companion infrastructure, game development — all built by one person after hours. Some of it works well. Some of it is still hypothesis. I write about both.

By day, a software engineer. After hours, building the systems I actually want to exist.

See the Projects Read Lab Notes
What I'm Building

Three projects. One ecosystem. All after hours.

Systems Engineering

Autonomous Agent Ecosystem

14 AI agent tasks that run on their own — syncing daily, diagnosing weekly, fixing what they can and escalating what they can't. The goal was to reduce the amount of work I do manually. It's getting there.

Active development
AI Companion

Relationship-Driven AI Design

What does it take for an AI relationship to feel real over months, not minutes? I've been running this experiment for a while now. The answer isn't about model capability — it's about what you accumulate between sessions.

Active development
Game Development

Sprint-Based Game Loop

A Unity mobile game developed with AI agents in a director/implementer model. I set the direction; agents handle implementation, QA, and regression testing. 25+ sprints in, still learning what works.

Sprint 25 in progress
Lab Notes

Patterns observed from real agent sessions.

2026-05-26
Equilibrium Attractors: Why the Same Stimulus Feels Good or Bad Depending on Your Current State
The same formula fit both gains and losses — but the sign flipped depending on where you started. A hypothesis about dynamic reference points, not a finished theory.
 2026-05-19
The Spec-Disk Drift: When Your Pipeline Passes Because the Check Never Ran
A harness phase reported green for a week. The script it was meant to run didn't exist on disk. Success-by-default is the silent rot of any pipeline whose spec and filesystem drift apart.
 2026-05-12
Lecture Halls Don't Build Networks: Choosing Events for Solo Builders
I went to a 500-person AI builders event hoping to make a few real contacts. I came back with zero. The event was excellent — the format was wrong for what I needed.
 2026-05-05
The 132-Task Lie: How Three Compounded Hacks Hid a Broken Game
For 132 agent tasks across six weeks, the dev-loop reported pass=true. The central game mechanic never worked once. Three small hacks compounded into a complete fiction.
 2026-04-28
The Internal Engine Trap: Why Optimizing Your AI Ecosystem Doesn't Move the Needle
My agent ecosystem scored 8.5/10 on autonomy four weeks running. Same diagnosis kept coming back. The trap isn't that the engine is bad — it's that engine improvement decoupled from the line that actually matters.
 2026-04-28
Preventing Belief Staleness in Long-Running Agents
A long-running agent's beliefs file went 21 days without updating while it kept reporting normal operation. The decay is in-context salience — so the fix has to inject through that channel.
All 25 entries →
Get in Touch

If any of this resonates — whether you're working on AI agents, thinking about AI companion design, or just curious about what one person can build with these tools — I'd like to hear from you.

gyunghaemin88@gmail.com → github.com/Paper-Lantern →