Unguided coding agents pass only 52% of web checks; Modern Web Guidance lifts it to 91%

rseroter · x · 2026-11-21

Chrome for Developers shipped Modern Web Guidance 1.0, a set of skills embedding web platform expertise, best practices, and browser compatibility data into coding agents. Philip Walton addresses the common question "don't frontier models already follow web best practices?" — testing says no. In their eval suite, the same prompts given to unguided agents pass on average only 52% of checks, while agents with Modern Web Guidance installed reach 91%. The post also catalogs the top 10 things AI gets wrong in web development. Takeaway: don't waste tokens adding skills for things LLMs already do well, but web best practices are exactly where explicit guidance pays off.

Original post →

More from coding & agent

coding & agent channel →