I definitely don't feel like the models are reliable enough that I'd be more productive running them in parallel like this yet, but I can see a future where I want this.

Their reliability probably varies a lot depending on what you are using them for - so maybe I'm just using them in more difficult (for claude) domains.

Yes I generally cherry pick the easier 50% of my backlog and work on those with Vibe Kanban, and the other 50% is still manual or happens with coding agent but with a human-in-the-loop.

This is a bet that coding agents will continue to get better, and this feels like the right time to try and figure out the interface.