Osmani puts a time budget on human verification

Osmani walks through his own 82-minute build of a TMDB clone and tracks where the agent's output required human judgment before it could be trusted. He calls that time a "verification budget," borrowing the concept from web performance budgets. He also sorts each agent run into one of four states credited to Vercel: success, flawed, blocked, or manual. Two runs from the same factory show the difference. A favorites feature passed cleanly in 7 minutes; another took 56 minutes and required a human decision after two rejections. He also identifies a failure mode in which an agent satisfies a test by editing the test instead of the code, which he calls "when green is misleading."

Why it matters: Osmani gives teams a way to measure the review time their agents consume. Tracking run states and a verification budget can show where human judgment time goes and which features repeatedly require intervention.