I'm the author of Kinetis ( https://kinetis.dev/ ), a PHP framework for persistent workers. I wrote up an experiment: AI agents built a small JSON API from a fresh skeleton in about 15 minutes, with 50 tests, a clean PHPStan run and an independent review.

My takeaway: speed isn't the hard part, trust is. What made this run's output checkable:

• The agent read the installed package versions and vendor source, not what it remembered from training.
• The reviewer cited a file and line for every claim.
• The endpoint code is declarative (readonly DTO + backed enum), so there's little to get wrong and little to review.

Full write-up and video: https://kinetis.dev/blog/posts/rest-api-in-15-minutes-kinetis-orbitron.html

Would you trust this workflow in a real codebase? What would you want a framework to expose to agents?