Agentic engineering is client work. Vibe coding is an experiment.

Andrej Karpathy coined the term on 2 February 2025 in a post on X: “There's a new kind of coding I call ‘vibe coding’, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.” In the same post he called it “not too bad for throwaway weekend projects”.

A few weeks later Simon Willison pinned the term down: vibe coding is “building software with an LLM without reviewing the code it writes”. If you review the code, test it and can explain it to someone else, that is software development.

That is the useful definition. Vibe coding is not a verdict on code written by AI. It describes a decision not to look at it.

One year later, on 4 February 2026, Karpathy looked back. Programming with LLM agents, he wrote, was becoming a default workflow for professionals, “except with more oversight and scrutiny”. His preferred name for it: agentic engineering.

“Agentic”, because “you are not writing the code directly 99% of the time, you are orchestrating agents who do and acting as oversight”. “Engineering”, because “there is an art & science and expertise to it” that you can learn.

In practice, agents write code, run tests and propose changes. You set the goal, split the work into pieces an agent can finish, read the diff, decide what ships and answer for it.

A client does not buy code. A client buys the assurance that it works, that it is safe and that someone stands behind it. None of that moves to the agent.

Quality comes from review, not from the model. An agent that writes a hundred lines in a minute also writes a hundred lines someone has to check.

Responsibility stays with whoever signs the invoice. “The agent did it” is not an answer a client accepts.

Traceability is the new part. When an agent ran for an hour on a client's project, the client should be able to see what it did, not only that it was billed.

Hours were a fair proxy as long as people typed the code. With agents, working time and effort drift apart.

Five minutes of prompting start an hour of agent work, followed by fifteen minutes of review. Several projects run at the same time. And whoever bills by the hour earns less exactly when they deliver more in less time.

That does not make effort irrelevant. The less you bill by the hour, the more you need to know what a project really took: at a fixed price, to see whether it paid off; on time and materials, to pass the effort on completely and in a way your client can follow.

• Effort: your working time, the agents' active runtime and the tokens they used, per project.
• Progress: what each session was about and which commits came out of it.
• Value: what the project was worth against what it took, whether billed by effort or at a fixed price.

You can sell vibe coding as what it is: a prototype, a mock-up, an experiment whose code nobody will maintain. Say so in the offer.

Software a client will run, extend and depend on is a different product. That is agentic engineering, and it is billed and documented like engineering.