I haven't tried the claude-on-rails gem yet. Even with good models, i think that it can become too easy to just allow LLMs/Agents to generate a bunch of code/prs and quality will suffer. I'm sure that there are some ways that this can be mitigated, but it does seem like this puts on a Team Manger hat instead of a Mentor hat when it comes to the agents. 

The computer I'm using on this episode is a M3 Ultra with 256GB of unified memory.