At 40:30 he doesn't use plan mode and believes that he is set up the system in such a way that it should be able to successfully complete tickets without plan mode. He doesn't read the plans.
...
not using plan mode
...
On reflection, I realize that I read plans because I don't trust two things: That the model will understand my intent, or that it will translate my intent into a structure that I will agree with. I realize that both of these could be addressed in a different way. Some people have taken to using a agent skill -- grill me skill to ensure that the model understands their intent. They suggest that they are getting much better results because of it. Another way of ensuring that the model will understand your intent is for your intent, to be clearly described at a higher level in the practice - context right in the repo. And secondly, if I have carefully specified architectural guidelines and guard rails, and have a very good testing infrastructure, then I should be less concerned that the model will structure things according to my desires.
...
I've been hearing about people moving away from plan mode, but I had a loose mental model that plan mode sort of forced the agent to think more thoroughly before it would actually implement, leading to better results. It's pretty clear that if you are not forcing plan mode to facilitate a manual review that simply using it to force the model to thank harder as part of a multi step process is an inductive bias that will be bitter lesson'd soon enough, if not already.