Everyone asks how fast you can build with AI.
Nobody asks how many times you send it back.
I just shipped OneDime for the PayPal AI Hackathon with an AI team. Claude as lead architect and reviewer. Codex writing the code. Gemini for art. Me in the middle of all of it, orchestrating.
Is it fast? Crazy fast. Four weeks. 312 tests. A live beta, a public repo, a 3 minute demo video, and a manga intro.
Is it easy? Nope lol.
What the back and forth actually looks like
- I asked for one file to be updated. The AI updated the wrong one. Undo, explain again, redo.
- The video motion pulsed on every beat. Looked great to the AI. Felt irritating to me. Switched it to a slow breathe.
- The voiceover said "No bots" while the screen showed... robots. Nobody caught it until I watched it with fresh eyes.
- The AI-drawn manga panels snuck a 7-Eleven logo onto a window and a designer monogram onto a hoodie. Had to scrub both.
- I told Codex "never push to the public repo." Then the public repo became the real repo. Codex stopped and asked. It was right to.
- One period at the end of a link broke the beta link an hour before submitting.
Every one of those got caught by a review gate or a human looking closely.
What I learned
Speed is the easy part. Judgment is the job.
AI suggests. The server decides. And I decide what ships.
You write the spec. You read the review. You catch what the agents miss. You send it back when it's not right. Then you do it again tomorrow.
It's still work. It's just a different kind of work... and I'd do it again.
Demo: https://www.youtube.com/watch?v=BbFXWSLfNaU Code: https://github.com/BCaris-RN/OneDime
Thanks for reading. If you're building with AI agents, what's the hiccup that surprised you most? Drop it in the comments.