Build 89% accurate AI agents still fail when users face complex pr