I pulled 30 days of Claude Code session data — 223 sessions, 38,000+ prompts, 32 projects, 1.2 GB of conversation. The numbers tell a story about what AI-native development actually looks like when nobody's watching.
Someone made the argument that most developers were never really engineering — they were sourcing solutions from Stack Overflow and Reddit, and AI just swaps the supplier. It's an uncomfortable take. It's also not entirely wrong.
I rebuilt a database table and it silently deleted two columns I needed. No error. Every test for the thing I was changing passed. Here's what actually caught it.
Three bugs that would have shipped clean. None of them threw an error. The only thing that caught them was a test written to fail first, against the real thing.
There's a file in my project whose only job is to stop me from making one mistake. I made it three times in a row. The difference between a rule and a wall.
Thirteen features shipped green. The designs were all wrong. The system made being wrong cheap—but only on one side of the work.
I once mapped the civil war between human interfaces and machine-native ones at the scale of companies. Then I fought both sides of it in a single afternoon, in a single pipeline — and a human had to close the seam.
I went looking for other cooks and found closed kitchens — not hostile, just complete. The Bear turned out to be the show that named what's actually missing in agent-assisted work, and what it might cost me to build the alternative.
I asked an agent a simple product question and watched it spend three thousand words rediscovering things it had no way to trust. The fix wasn't a better map. It was noticing which artifacts in a codebase can lie to you — and which one can't.
My lawyer told me to use ChatGPT instead of paying him. A colleague said the same about his accountant. The cleanest signal for which professions get compressed isn't the headcount data — it's when the practitioner himself tells you he's optional.