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 asked the new Siri what it could infer from a week of texts. The answer was not a better chatbot. It was inference over a graph no OAuth connector gets.
A useful design reference may come from a completely different product. The work is identifying the decision it helps with, then telling the agent which part should transfer.
Using AI takes practice, but a product that needs an expert to catch its mistakes should say so before someone relies on it.
I asked an AI agent to audit my app for the person it was built for. It sent five sub-agents to check documents and none of them looked at the screen.
A small inspection of one runaway session became repeated parallel work before it had earned that expansion. The 119-of-118 report was the alarm.
My machine ran out of resources, so I handed an agent four screenshots and typed one word: fix. It did a good job. A day later, twenty-two connector processes were still running, every one of them marked disabled in config. Nothing had failed — and that turns out to be the interesting part.
Last October I published a post arguing that agent safety is an architecture problem: the model proposes, hard-coded code disposes, a human approves. Then I built exactly that for a youth sports club's website and used it for an hour. I found seven things wrong with it. None of them were in the part I had been afraid of.
A local model run took my Mac down at 10:55:33. Recovering the research cost about twenty minutes. Writing down the wrong reason for the crash cost more, and the file that proves it wrong ships with the operating system.