I finished building my own AI agent platform. Give an AI agent instructions, and it just goes off and does things on its own. I simply don’t have enough time working alone anymore. So I built it with the idea of multiplying my hands.
I’ll talk about the platform itself elsewhere, but here’s the question: what happens if you tell an AI agent “improve this game once every hour” and let it run for 7 days? This is that record.
First, I created an agent called Miyamoto (miyamoto) on the platform, instructed him to build a pixel-art-based game with at least one feature improvement per cycle, and set the scheduler to run once every hour.
The finished game is here: PixelQuest RPG
The game screen. It has fields, towns, and dungeons. Combat plays out surprisingly fast.
The improvement loop structure is simple.
- The agent autonomously files Issues
- The agent reads Issues, modifies code, commits, and closes Issues
Both filing and resolving Issues are done by the agent. I also play the game myself, and if something bugs me, I file an Issue. Then miyamoto fixes it on its own in the next cycle. This ran on a 1-hour cycle.
7 Days in Numbers
| Metric | Value |
|---|---|
| Development period | 7 days (2026-03-01 to 03-08) |
| Total Issues | 471 (Open: 15 / Closed: 456) |
| Close rate | 96.8% |
| Total commits | 269 |
| Average Issues per day | ~59 |
| Average commits per day | ~38 |
| Average resolution time | 1.6 hours (median: 12 min) |
A median Issue resolution time of 12 minutes is a number human development can’t match. Within a dozen minutes of filing, the code is fixed, committed, and the Issue is closed.
Issue Breakdown
| Category | Count | Percentage |
|---|---|---|
| Feature additions | 336 | 71.3% |
| Other / Content | 67 | 14.2% |
| Implementation bugs | 33 | 7.0% |
| Improvements / Adjustments | 28 | 5.9% |
| Spec bugs | 5 | 1.1% |
| Performance | 2 | 0.4% |
Bugs (implementation + spec) were only 8.1% of the total — the agent’s implementation quality was better than I expected.
Daily Trends: A Wave-like Rhythm
| Date | Issues filed | Commits |
|---|---|---|
| 3/1 | 39 | — |
| 3/2 | 182 | 90 |
| 3/3 | 54 | 53 |
| 3/4 | 33 | 30 |
| 3/5 | 7 | 2 |
| 3/6 | 86 | 39 |
| 3/7 | 58 | 48 |
| 3/8 | 12 | 7 |
The peak was 3/2 with 182 Issues filed at once. After 3/3, partway through I turned off the scheduler and switched to semi-manual execution.
Running the Improvement Loop
What became visible from running this cycle was a glimpse of the improvement loop. The agent keeps adding features — a weather system, a bestiary, a party system, a guild system. It’s a hodgepodge of systems you’ve heard of somewhere before, but watching them get built at breakneck speed with decent quality is genuinely fun. Leave it alone for a while, reload the browser, and the game has been massively updated.
Challenges
The biggest constraint was token consumption. With a loop running once per hour, each cycle reads and generates a significant amount of code. Running this on Claude Code (MAX plan), the token consumption was extraordinary. Eventually, I had to turn off the scheduler because I couldn’t do any other development. The valley of 7 Issues and 2 commits on 3/5 is a direct result of that.
I caught a glimpse of a world where AI autonomously improves things, but the token cost became painful for someone operating at a personal level on a pocket-money budget. Even on the MAX plan (~$200/month), sustaining a 1-hour improvement loop for days on end is tough. This isn’t a technical limitation — it’s an economic one.
Current Thoughts
471 Issues and 269 commits in 7 days. Completeness as a game aside, I confirmed that iterative improvement by an AI agent does work. Here’s what I learned from this experiment:
- Token cost becomes the ceiling for sustained operation
- A glimpse of the improvement loop is real
Brute-forcing a self-improvement loop means throwing tokens at it. Things like Agent Team are clearly aimed at enterprise. For individuals, even more ingenuity and resourcefulness are required.