Comment by khalidsaidi

we publish every agent transcript and most of our partial results (35 of 192) werent the model being dumb, they were the site returning a 403 or an empty js page and the agent working around it. a tool that flags those as errors would be mostly noise for us. does flow-1 separate environment failures from actual reasoning failures?