Why coding agents are quietly reshaping frontend development

Nolan Lawson, a longtime web performance engineer, wrote a sharp piece on what AI agents are doing to frontend development. His argument is not that agents write bad code. It is that they change which knowledge is worth having, and the incentives are shifting in ways that are easy to miss.

Three shifts stand out. First, frontend code is comparatively cheap to replace, so teams ship agent-written code with less review than they would give a database migration. Second, agent familiarity is starting to beat developer ergonomics when teams choose tools. He points to Cursor and Viget moving off technically stronger frameworks like Solid and Lit and onto React, simply because agents were trained on far more React and produce better results with it. Third, browser vendors are leaning toward new raw capabilities over friendlier syntax, on the assumption that agents will handle the complexity anyway.

Lawson is not fatalistic. He tested Claude Sonnet on a genuine performance brain-teaser about optimizing Chrome trace rendering and got a careful answer covering selector complexity, invalidation scope, and how to measure it. His response is to adapt: teach agents better architectural defaults, build sites that agents can navigate, and expect consulting work untangling AI-generated code that quietly became load-bearing.

Why it matters

If you pick frontend frameworks, "which one do agents know best" is now a real selection criterion, not a joke, and it may pull your team toward React regardless of technical merit. The deeper risk is skill erosion: if reviews get lighter because frontend feels disposable, the load-bearing bugs still land in production.