My experience with heavy AI-agentic development:
- I have a Claude Code Max 20x and ChatGPT Pro 20x ($200 each)
- I prompt them directly (no work queues yet)
- I average four threads in parallel
- Around 2/3 of my time is spent QAing the agents' work. 1/3 is invested in ideation.
- Codex and Claude are very similar. The outcomes rely most on the harnesses I give them.
- I develop with both of them in the cloud and locally.
- Claude's UX is better.
- I'm tokenmaxxing by the sixth day of each week.
Of course, all this will be replaced when switching to automated development with queues and automated testing. The latter will take most of the time. Fortunately, testing won't require the most expensive models.