OK but if the verification loop really makes the agents MUCH more useful, then this usefulness difference can be used as a training signal to improve the agents themselves. So this means the current capabilities levels are certainly not going to remain for very long (which is also what I expect but I would like to point out it's also supported by this)