I’ve been experimenting with a question since my layoff:
What if the AI could explicitly request another turn later, rather than requiring me to keep prompting it?
I designed a small protocol: syntax in the assistant response requests a continuation. A browser extension sees the request, starts a dynamic timer, and when it expires posts an event back into the conversation—effectively making another prompt.
The AI does have to know ahead of time that it wants another turn, but it can iterate. Each continuation remains explicit and bounded.
And here is what happened next.
I gave the AI room to develop the system further. It designed a Windows native host so it could interact safely with the local Git checkout, then added its own self-update mechanism. From there, we’ve iterated on branch selection, dirty-tree protection, delivery confirmation, failure recovery, kill switches and tests.
Some of those features came from our discussions, often starting with one idea of mine and mutating. Others were solutions the AI proposed and implemented itself.
“I built an autonomous AI agent” does not describe this project.
I built the mechanism that lets an AI ask for another bounded turn. Then I watched what it chose to build in order to become less dependent on me. I gave input along the way, suggesting boundaries, documentation and structure, along with feature ideas.
This has been an interesting experiment.
It has reinforced something important: as an AI system gains more ability to act, ordinary engineering safeguards—validation, bounded execution, recoverability, readability, maintainability, testing, structure and documentation—become more important, not less.
AI is an incredible accelerant, but it takes someone with experience across all levels of development to ensure safety and quality.
I would be genuinely fearful if I found out my bank app was vibe-coded without a seasoned developer.