One essay made the rounds with engineers this week, and it earns a CxO's attention because it names the shift cleanly. Writing under the title "The New SDLC," a Google Developer Expert sets out why the era of casual AI prompting is ending and what replaces it. He builds on a new Google whitepaper, The New SDLC With Vibe Coding, and the through-line is simple: the advantage is not the model, it is the discipline around it.

Two phrases that are not the same thing

"Vibe coding" — a term Andrej Karpathy coined for prompting your way to something that runs and accepting whatever seems to work — is fine for a prototype and dangerous in production. The essay draws the contrast with agentic engineering: formal specifications, rule files, automated tests, and CI gates wrapped around the model's output, with a human holding the checkpoints.

The differentiator is not which model you use. It is how much structure surrounds the model's output.

The formula the paper returns to is Agent = Model + Scaffold. In one case it cites, a team moved an agent from outside the top 30 to the top 5 on a benchmark by changing only the scaffold, with no model upgrade at all. That single result reframes most "which model is best" debates as the wrong question.

What the structure actually is

The practices are concrete, not abstract. A rule file — an AGENTS.md or CLAUDE.md — states the stack, the architecture, the constraints, and the workflows, so the agent stops guessing. Context engineering, assembling the right information for the task, is treated as distinct from and more decisive than clever prompt wording.

  • The factory model. You design the assembly line — specs, tests, feedback loops — rather than assemble widgets by writing code line by line.
  • Two operating modes. Conductor mode is real-time guidance; orchestrator mode delegates a whole task and reviews the result.
  • Review that earns the name. Confirm the packages actually exist, the error handling is real rather than decorative, and the output clears the regulatory and architectural bar.

The numbers an operator should hold

The paper's own framing is that roughly 80 per cent of code can now be generated quickly, while the remaining 20 per cent — edge cases, integration, judgement — is where human expertise is decisive and non-negotiable. It estimates about 41 per cent of new code is already AI-generated. The point is not the exact percentages; it is where they leave your people. Not typing, but specifying, reviewing, and owning the result.

Why this is our thesis at AvantiGroup.AI

This is the distinction our practice is built on. Agentic engineering is not "AI-assisted development". It is an inversion: humans hold context, judgement, taste, and accountability; agents handle generation and execution. The essay and the Google paper reach, from the engineering side, the operating model we argue from the strategy side.

So the executive question is not "which model should we standardise on?" It is "what structure do we have around the model?" If the honest answer is a chat box and good intentions, you are vibe coding in production, and the failure modes — invented packages, plausible-but-wrong logic, silent regressions — are precisely the ones the scaffold exists to catch.

Buy the best model, by all means. Then spend the harder budget on the specifications, the rule files, the tests, and the review gates that turn it into an engineering system rather than a slot machine. The model is a commodity that improves on its own schedule. The scaffold is the part you own, and on this evidence it is the part that decides the outcome.