If you're looking for a "vibe-coder guy" — here's what I think about that as of March 2026.

You can absolutely set up a pipeline today where a fleet of AI agents automatically picks up Sentry errors, writes fixes, and ships code to production without a human touching it. The tooling is there — Claude Code, ChatGPT Codex, Cursor IDE, agentic workflows. I use all of these daily for different tasks. This very site was built with them.

But I still believe — in March 2026 — that you need a human in the loop. Someone who reviews the code and catches architectural problems before they compound. That person should:

• Know how to work with agents — this multiplies how many problems you can find and fix.
• Have system design knowledge — to catch when an LLM makes a bad architectural call.
• Know the language and frameworks deeply — the subtleties that prevent performance and security issues.

I've spent time digging into how neural networks actually train and work under the hood. I think everyone should — not to become an ML engineer, but to stop treating it like a magic box. Understanding what's inside helps you work with it better. It's like knowing how a manual gearbox works — it genuinely helps you drive.

If you haven't yet, I highly recommend this video — some details may be outdated by now, but the fundamentals haven't changed.