Andrej Karpathy defined "vibe coding" a year ago. He just killed the term.
The replacement? Agentic engineering.
We are no longer just writing code. We are orchestrating agents and reviewing their output. Here's what has actually changed on the ground:
-
Prompting is the easy part: Architectural precision matters more than clever prompts. Define your architectural decision first and you will be good.
-
Code review just got hard: The agent writes fast. Catching bad error handling, missed edge cases, security gaps? That's on you. Have you thought about the user who is going to do C while the AI only planned A, B because of its token limit?
-
Architecture hasn't changed at all: Which patterns to use, how components fail, what breaks at scale. AI doesn't help here.
The skill gap isn't prompting. It's the same gap it's always been: can you architect, can you review and validate properly, do you know the failure modes.
If you're shipping with AI agents daily, what surprised you most about what changed and what didn't? #agenticdevelopment