What they did
Before launching a 20-product shop site that their local agent had built, Nazoor Ahmed ran two independent UX and SEO audits from the same brief: every page at phone and desktop width in light and dark mode, every visitor flow, a dark-mode verdict, an SEO review, before-and-after screenshots for each top issue, and no changes to site files. One audit was their own agent engine running a local 27B model with browser tools, overnight. The other was Claude, which handed the brief to an expert sub-agent with its own browser. Claude then acted as referee, confirming each finding in code or the browser and merging both reports.
How it went
Claude's auditor took about 40 minutes and found 28 of the 32 merged fixes and 4 of 5 launch blockers, including a placeholder order phone number and a wrong price on larger sizes. The local agent took about 14 hours, 475 model calls and 430 tool actions, found 12 fixes, but caught a price filter that emptied the shop, which Claude missed. Its review loop stalled for 45 minutes on proving no file changed, and two of its screenshot pairs showed nothing.
Worth knowing
The local agent is thorough and literal: it tests what it is told and misses what a person would notice fast. Checks that only confirm a file exists do not prove a screenshot shows the fix. The audit ran on a copy of the site so the no-changes rule could be verified by comparing folders.