Building an AI development harness that learns
A self-learning loop belongs at the center of an AI development harness. 🔁

When an agent discovers something that materially improves the work, that learning should become available to other projects and future tasks automatically.
That’s what I’m building into my development workflow.
At work, I’ve created a shared AI harness for our team. I also maintain one for Lucky Dreamer, my solo game studio.
The harness brings together the instructions, reusable skills, tools, accumulated evidence and verification workflows that guide how my agents work.
My studio setup shares the same core across Codex and Claude, locally and in the cloud. Gemini is another option I’m evaluating.
🧠 The learning loop works like this:
- Identify a discovery that could benefit other tasks or projects.
- Validate it and capture the evidence, conditions and tradeoffs.
- Contribute a focused, tested improvement to the shared skills, tools or knowledge for review.
- Automatically retrieve relevant, approved lessons during future work.
That could be a better debugging approach, a performance improvement, a stronger testing strategy or a useful failed experiment.
Later evidence must also be able to challenge earlier advice. A fix that helps one environment can hurt another.
🎮 My studio is a practical testbed. If I’m working on 20 games, a valuable discovery in one should inform work on the other 19 and the games I build next.
I shouldn’t have to remember every discovery and manually copy the best parts between projects.
This is where I’m spending a lot of my time: making the development system improve alongside the software it produces.
My personal harness combines ideas from pstack and other open source agent workflows with a year of my own experiments, mistakes and discoveries.
🚀 I’m considering open sourcing the Lucky Dreamer harness. If you’d find it useful, let me know. That might be the nudge to publish it.
#AIAgents #AIEngineering #DeveloperTools #BuildInPublic