# Vibe Coding and Agentic Engineering Are Getting Closer Than I'd Like

Simon Willison reflects on a recent podcast appearance during which he realised that the boundary he had drawn between *vibe coding* (irresponsible, unreviewed AI code generation) and *agentic engineering* (professional, code-reviewed AI-assisted development) had blurred in his own practice.

His earlier framing was clean: vibe coding is fine for throwaway scripts, never for production; agentic engineering is the responsible variant where a competent engineer reviews every line. As coding agents like Claude Code became more reliable, he stopped reviewing every line, even on code he ships. He admits this directly: "I'm not reviewing that code. And now I've got that feeling of guilt."

He compares it to how engineering teams treat services from other departments — somewhat as black boxes until something goes wrong. But the analogy breaks down at accountability: "Claude Code does not have a professional reputation! It can't take accountability." A human engineer's reputation is the implicit collateral that lets you skip review.

A second observation: AI-generated repositories now arrive looking polished — comprehensive READMEs, full tests, clean directory layouts. The visible markers that used to signal effort no longer do. He suggests the new heuristic is whether someone has actually used the thing: "I want somebody to have *used* the thing."

A third: throughput has gone from ~200 to ~2000 lines per day, which moves the bottleneck both upstream (design and specification) and downstream (testing, verification, integration).

He closes by reaffirming that programming as a craft is "ferociously difficult," that AI is an amplifier of existing expertise rather than a replacement, and that experienced human software engineers will continue to have value — but does not deny the discomfort that accountability has become detached from the artifact-producer.