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

Simon Willison's 2026-05-06 post following a podcast conversation. Earlier he had drawn a sharp line: vibe coding is unreviewed AI code generation, fine for throwaway scripts and dangerous for production; agentic engineering is the responsible variant where a competent engineer reviews every line. The post is him watching that line dissolve in his own day-to-day, and naming the discomfort.

The load-bearing admission is unambiguous: "I'm not reviewing that code. And now I've got that feeling of guilt." Production code, written by Claude Code, shipped without line-by-line review. He doesn't dress this up as a new methodology — he labels it correctly as drift, and reports the guilt rather than rationalising it.

Trust without accountability

His attempt to make peace with the drift is to compare AI coding agents to services run by other engineering teams: you trust them, you don't audit each commit, you intervene when something breaks. The analogy admits its own break: "Claude Code does not have a professional reputation! It can't take accountability." The reason you can skip the audit on a peer team is that there's a human at the other end whose career is implicit collateral. With an agent, the collateral is missing — the trust shape is the same, the trust foundation isn't.

This is the demand-side counterpart to Faye's structural argument: the orchestrator role assumes a stable human supervisor; Willison's post is the supervisor reporting that he stopped supervising and felt bad about it. The post adds Willison's name to the empirical leg of skill-atrophy-supervision-paradox — already cited there for losing his "firm mental model of what applications can do," now also as a self-reported review-skipping incident.

Polished output is no longer a quality signal

A second observation worth keeping. Generated repositories now arrive with full READMEs, comprehensive tests, and clean directory layouts; the visible markers that used to mean "someone cared" are now table stakes. Willison's new heuristic is that the only durable signal is whether somebody actually used the thing. "I want somebody to have used the thing." This is the same ground features-to-steal-from-npmx covers from a different direction (download counts as anti-typosquatting signal) and that specsmaxxing gestures at (the spec is what survives because the artifact got cheap). When the artifact is cheap, attention moves to use, to specs, to people.

Bottlenecks moving outward

Throughput has gone from ~200 to ~2000 lines/day. Willison flags the cascading effect: the bottleneck moves both upstream (what should we build, against what spec) and downstream (testing, verification, integration). Same pattern as no-silver-bullet-llms — Brooks' essential difficulty is unchanged, so accelerating accidental difficulty just relocates the work. specsmaxxing is the upstream half of this, agentic-coding-fatigue's 4-5 hour ceiling is the downstream half.

Where this fits

Willison sits awkwardly between the cluster's poles. Unlike i-am-an-ai-hater (Moser's refusal) or cult-of-vibe-coding (Cohen's read-every-line discipline), Willison reports the drift without prescribing a return. Unlike agentic-coding-is-a-trap (structural argument) or no-silver-bullet-llms (theoretical argument), he isn't arguing — he's confessing. That makes it useful as testimony rather than analysis: a senior practitioner with reputation and audience visibility, naming a behaviour change he has not chosen to defend.

The closing reaffirmation — programming is "ferociously difficult," AI is an amplifier, experienced humans still have value — is honest but does not resolve the discomfort. The post leaves the contradiction open: he believes the supervisor role matters, he is failing to perform it, he is going to keep shipping anyway.

Coverage in the wiki

Cross-references: agentic-coding-is-a-trap, skill-atrophy-supervision-paradox (Willison is direct evidence for the "lost mental model" leg), cult-of-vibe-coding, agentic-coding-fatigue, peril-of-laziness-lost, building-syntaqlite-ai, no-silver-bullet-llms, specsmaxxing, anti-llm-discourse (Willison sits outside the five sub-genres — neither hater nor evangelist).