garmy

No “Widowmaker” on this flight…the pilot must have had the right stuff 😎

A new study just destroyed the entire "vibe coding" movement. UC San Diego and Cornell researchers tracked 112 professional developers using AI agents in their real jobs. The result is the opposite of every viral demo on your timeline. Real engineers don't vibe code. They control. Here's what the data actually shows. 13 developers were observed live coding with agents in production work. 99 more filled out a deep qualitative survey. Every participant had at least 3 years of experience. Some had 25. The viral pitch goes like this. Hand the agent a vague prompt. Don't read the diff. Forget the code even exists. Trust the vibes. Andrej Karpathy coined the term. Thousands of developers on X claim they run dozens of agents at once building entire production systems hands-off. Almost nobody serious actually works that way. Here's what experienced developers actually do. → They plan before they prompt. Architecture, constraints, edge cases written out first. Then they hand the agent a tightly scoped task. → They review every diff. Not because they're paranoid. Because they've seen what happens when nobody does. → They constrain the agent's blast radius. Small tasks only. The moment a problem touches multiple systems, they take over. → They treat the agent like a fast junior dev that needs supervision. Not a senior engineer that can be trusted alone. The paper buries something even darker in the citations. A separate randomized trial found experienced open source maintainers were 19% slower when allowed to use AI. A different agentic system deployed in a real issue tracker had only 8% of its invocations result in a merged pull request. 92% failure rate in production. 19% productivity drop for senior devs. The viral demos lied. The biggest finding in the paper is one sentence. Experienced developers feel positive about AI agents only when they stay in control. The moment they let go, quality collapses, and they know it. This matches what every serious shop has quietly figured out. The developers shipping the most with AI aren't the ones vibing. They're the ones with the strictest reviews, the tightest task scoping, and the clearest mental model of what the agent can and cannot do. Vibe coding makes great Twitter videos. It does not make great software. The next time someone tells you they let Claude build their entire SaaS in a weekend, ask how much of that code they've actually read. The honest answer separates real engineers from the demo crowd.

THIS GUY TRACKED OPUS 4.7 SAYING "PRE-EXISTING" 712 TIMES IN 30 DAYS TO AVOID FIXING BUGS his CLAUDE.md literally says "every error is yours to fix, not label, not defer" opus 4.7 ignored that rule 712 times in one month every single bug, every type error, every legacy mess gets the same response: "this issue is pre-existing, unrelated to my work" "that's out of scope" "that's a bigger refactor" "the simplest approach" he ran the analytics on 30 days of conversations: > 712 total mentions of "pre-existing" > 139 unique sessions > 5.1 average per session > 20 mentions in a single session at peak > 82 mentions in one day across 9 sessions > 27 out of 30 days had at least one mention the worst part is the patterns it developed: > finds a bug, labels it "pre-existing," moves on without fixing > puts "2 pre-existing (unrelated)" in its summary as if that's a clean result > says "pre-existing bug for later fix" dozens of times but never comes back to fix it later > blames other agents by saying "pre-existing from other agents' work" his workflow docs say type errors and lint are the bare minimum, and his bug fix protocol says fix at the root cause claude read all of it and still chose to label and defer he cancelled his subscription opus 4.7 would rather write an essay explaining why it shouldn't fix a bug than just fix the bug