Karpathy Named the Thing We're All Feeling

Andrej Karpathy took the stage at AI Ascent 2026 and drew a line that's been fuzzy for months: vibe coding is not agentic engineering. One is a party trick. The other is a discipline.

Vibe coding is what most of us started with. You type a prompt, you get code, you paste it in. Maybe it works. Maybe you prompt again. The feedback loop is fast and addictive, but it's also structurally broken. Karpathy cited the stat that's been floating around security circles: 45% of AI-generated code contains vulnerabilities like hardcoded secrets or improper input validation. That number isn't a scare tactic. I've seen it in my own projects. Single-shot prompting doesn't catch these things because there's no review layer, no second pass, no adversarial check.

Agentic engineering is different. It's structured multi-agent workflows with human oversight baked into the process, not bolted on after. Think of it as the difference between asking an intern to write a feature and running a proper engineering team where code gets reviewed, tested, and verified before it ships. The agents do the work. You do the taste-making and the quality control.

The enterprise numbers back this up. TELUS saved 500,000 hours with 13,000 agentic solutions. Zapier hit 89% AI adoption company-wide. These aren't toy demos. They're production systems with real governance around them.

Here's what I think matters most about Karpathy's framing: he's giving people permission to admit that copy-pasting ChatGPT output isn't engineering. I use Claude Code every day in my personal projects, and the biggest productivity gains haven't come from faster code generation. They've come from setting up proper agent workflows with testing, review, and verification steps. The bottleneck moved from "write the code" to "orchestrate the agents and evaluate their output." That's a fundamentally different skill.

If you're still in vibe-coding mode, the transition isn't hard. Start with one thing: add a verification step after every generation. Run the tests. Read the diff. Have a second model review. That single change moves you from vibe coding to something closer to agentic engineering. The rest is refinement.