AI coding agents are the new rockstar developers — and they leave the same mess

Jesse Skinner draws a parallel between a familiar industry archetype — the brilliant ‘rockstar’ developer who rewrites a company’s architecture in clever, idiosyncratic code and then leaves — and the behavior of LLM coding agents. The rockstar’s legacy is a codebase nobody else can run, debug, or extend; Skinner, who specializes in rescuing such projects, argues that the defining trait of these developers is speed and cleverness prioritized over code that teammates can actually maintain.

LLMs, he contends, replicate this pattern at scale. Each new chat session is effectively a fresh rockstar with no memory of yesterday’s decisions, happily emitting thousands of lines that ignore the system’s existing conventions. Agents over-apply generic best practices, layer on ‘belt and suspenders’ complexity, and produce review feedback that raises the bar without improving understanding. The result is worse than a single rockstar’s mess: vibe-coded systems assembled across hundreds of disconnected contexts, with no unifying design at all — sometimes accruing technical debt that can never be repaid. Teams risk becoming dependent on LLMs just to comprehend the code LLMs wrote.

His prescription isn’t abstinence but discipline: keep humans leading the engineering, have models generate small, reviewable pieces, and slow down when you can no longer explain what the code does or why. Simplify until the architecture matches the actual problem, and accept that writing some code by hand — craftsmanship — is the one thing that can’t be outsourced to a machine.