The term that named a whole movement is already being retired by the person who coined it. In early 2025 Andrej Karpathy described vibe coding — building by describing intent to an AI and “giving in to the vibes”. By early 2026 he’d called his own coinage passé and proposed a successor: agentic engineering — AI agents that autonomously plan, write, test and ship code under structured human oversight.
It sounds like the problem just solved itself. The agents are more capable, they write their own tests, they run in loops. So surely the human-in-the-loop part fades away? It’s the opposite. The whole definition rests on three words most people skip: under human oversight.
More autonomy raises the stakes, it doesn’t remove them
Vibe coding produced a prototype you could eyeball. An agentic workflow produces more code, faster, with less of you in the loop — an agent can write a third of a production codebase in a week. That’s not less risk to review; it’s more surface area generated by something that still shares the blind spots that created the bugs.
An agent that plans, writes and “verifies” its own work is grading its own homework at speed. It will confidently ship a data model with no access rules, a secret in the client, an endpoint with no limits — and its own tests, written against its own assumptions, will pass. The faster the loop, the faster those reach production.
”Structured human oversight” is the product, not the footnote
The reason agentic engineering is the grown-up version of vibe coding is precisely that it puts a human checkpoint back into a process built for speed. The discipline isn’t “let the agent run” — it’s “let the agent run, then a senior reads it before real users do.”
That’s the modern setup we’ve described before: build fast, then a senior reads it before production. It doesn’t slow the agents down. It just means the one judgement call that matters — is this actually safe to put real users and real data behind? — is made by someone who isn’t the thing that wrote the code.
What the human checkpoint actually catches
The same boring, invisible failures that take down launches, now generated faster:
- A database an agent left readable by everyone.
- A credential an agent inlined to “make the example work”.
- A tool or API the agent gave itself broad access to, with no guardrail.
- Architecture decisions made silently, three loops ago, that no one reviewed.
None of these show up in the agent’s green test run. They show up when a person reads the code asking “what’s the worst this lets a real user do?”
Call it vibe coding or agentic engineering — the name keeps changing, the gap doesn’t. seniorgrade is the human read the workflow is supposed to include: a senior engineer reads what your agents shipped and tells you, in plain English, whether it’s safe to launch. The agents got faster. The reason to have a human look first got stronger.