Miguel Pires da Rosa

Someone builds a project management tool with Claude Code over a weekend. Ships it. Tweets "just replaced Jira." The app works. One user, happy path, localhost. Then two people edit the same record simultaneously, and the data is silently corrupted. They don't know what an optimistic lock is. They never needed to before. The prototype is maybe 1% of what makes software actually work. The other 99% is what you find after real users show up: race conditions, failed transactions, sessions expiring at the wrong moment, a payment webhook that fires twice and charges someone double. AI didn't cover any of that. It built exactly what you asked for. And the confidence is the worst part. "Just need to adjust a few things before we go live." The few things you need to adjust are the product. That's like laying a foundation and telling people you basically built the house. Vibe coding works. For personal tools, throwaway scripts, and prototypes you'll never put in front of paying users, it's genuinely fast and good enough. I use it. But there's a hard ceiling, and it shows up the moment the stakes get real. Agentic engineering is a different discipline. You're not prompting for code. You're decomposing problems, designing system boundaries, writing specs precise enough that the agent doesn't go sideways. You review everything it builds, because it will make mistakes that only look wrong if you know what correct looks like. You guide it. You catch what it misses. If you don't know what a distributed transaction is, the agent won't save you. It'll generate something broken with complete confidence, and you won't know until production. The hard part of software was never writing the first 200 lines. It never was.