I've been exploring a question that I don't think we've really answered yet: can an AI agent — working mostly autonomously, with limited human oversight — build large, complex software you'd actually maintain for the long run, not just throwaway snippets?
Out of that exploration came a method I've been calling BASE — Bounded-Autonomy Software Engineering. The short version: the ceiling on AI-generated code probably isn't the model, it's the system you put around it — how you bound the autonomy, where the human engagement lives, and how quality gets enforced mechanically instead of hoped for.
I finally wrote it up as a white paper. Still early, still evolving, and I'd rather share it now and be wrong in public than sit on it.
Mostly I'm curious who else is poking at this space. If you're working on agent orchestration, bounded autonomy, or getting AI to produce code that survives contact with a real codebase — I'd love to compare notes. What's working for you? What isn't?
First published on LinkedIn on 12 July 2026.
Subscribe by RSS to get these when they land.
Promise is built by AI agents directed by one maintainer. It got here on a single ~$200/month subscription; roughly every $250/month is one more agent working full time. Sponsor Promise