Web development veteran compares AI's impact on frontend to an asteroid impact

Developer Nolan Lawson published a piece on August 23 on his blog Read the Tea Leaves about what's happening to frontend development in the age of agents. The main metaphor made it straight into the title.

Lawson has been a known name in web development for a long time. He worked on browser performance for the Microsoft Edge team, contributed to the open source Lightning Web Components framework at Salesforce, worked at Squarespace, wrote a Mastodon client called Pinafore, and maintained the PouchDB database for several years.

The piece grew out of watching what's happening to his peers. Several well-known frontend educators, including Axel Rauschmayer, Salma Alam-Naylor, and Josh Comeau, have scaled back or shut down their work entirely. Others shifted from web development to talking about AI instead.

Lawson decided to test the idea himself. He grabbed his favorite browser comprehension test – a Chrome trace with expensive style recalculation that trips up even experienced developers – and handed it to Claude Sonnet.

The answer held up well. The model even brought up obscure details like the cost of duplicating styles across shadow trees. Lawson admits he now just dumps traces into Claude Code himself, and the approach has already paid off at his day job.

The post lists four trends working against investing in deep frontend knowledge:

• Frontend code is less risky to hand off to an unsupervised agent, since a database migration needs review and staging, while a React component can usually just ship as-is

• Developer ergonomics matter less as agent ergonomics take priority

• Standards will shift away from syntax convenience toward raw platform capability, since an agent doesn't care whether it writes three lines of CSS or one

• New syntax actually gets in the way, since it isn't baked into a model's weights and the agent has to be manually taught to use it

Lawson also points to Cursor and Viget, both of which moved their codebases from Solid and Lit to React. Cursor's reasoning was blunt – agents know React.

The details of Cursor migration are telling. It started as an experiment with long-running agents, took more than three weeks, and produced 266,000 added lines and 193,000 deleted lines. The result passed CI, but still needed careful review.

Cursor engineer Lauren later said the migration is essentially done, with the new agent window now 99% React. The official reasons given are performance and maintainability, since signal-based reactivity easily creates traps when dependencies grow unpredictably in large apps. She also noted that agents struggled to write clean Solid code, since everything ended up accidentally tracked.

Nolan Lawson puts it this way:

An asteroid just hit the earth, and we’re still surveying the wreckage.

There are some more optimistic scenarios in the piece too, though the author calls them shaky. Agents still need to be taught the bigger picture, since they love building React single-page apps in cases where Astro or Eleventy would do the job with half the code.

A second path involves building sites that work well for agents, which loops back to fundamentals like server-side rendering, accessibility, and load speed.

A third option looks like a niche for consultants. The sheer volume of AI-generated frontend code is already becoming business-critical, and slow, insecure, non-compliant sites won't get fixed just by asking an agent to "fix my site."

Lawson doubts the web in its current form will even survive the spread of agents. He'd rather find out the cost of a flight from Seattle to Paris through an agent than by clicking through a slow website, and the only things stopping that right now are bot blocking and the lack of MCP support.

Discussion around the post turned to homogenization. One commenter suggested calling it "leveling," another pushed back, saying leveling works more like a boot flattening any uniqueness. The counterargument landed just as bluntly – the world probably doesn't need forty web frameworks that all look alike.

Lawson does see one area holding up against replacement. Replying to a commenter, he noted that agents currently have bad taste, don't understand client requirements, can't reason through user scenarios, and on top of that have no eyes – so a move into design might be a smart career step for a frontend developer.

Asked about an inevitable bubble burst, he replied that the changes will be sharp, but the language models themselves aren't going anywhere.

As evidence, he pointed to open weights, noting that Kimi K3 already handles routine coding tasks just fine, and that models are gradually becoming capable of running on consumer hardware.