I would be interested in reading what tools are made available to the LLM, and how everything is wired together to form an effective analysis loop. It seems like this is a key ingredient here.
Indeed, these may be the last ones to be fired, as they can become efficient enough to do the jobs of everyone else one day.