I've started instructing Claude to request code reviews from other Claude instances and to iteratively improve its code before considering it complete. Given that longer interactions can degrade quality and introduce biases, I wanted to know if this method is too complicated or unnecessary. It took a bit of effort to set up the child Claudes correctly, managing permissions and avoiding timeouts, but it seems like it's functioning now. I do have concerns though—like not being able to monitor token usage or context data for the child instances, and while they're working, it appears the parent is stuck when it's actually just waiting. If there's a smarter tool out there that does this better, I'm all ears!
2 Answers
What settings are you using in your Claude.md regarding child management or multiple Claudes? Some specific configurations can help manage this more effectively.
You're not over-engineering it at all! This is how you maximize the effectiveness of these models. The scaling process discussed in the official documentation supports this kind of setup very well. Trust me, it’s a common practice among developers! You can even use Ctrl+R to view subagent steps if you ever want to check in on the process.
In ORCH mode, you can use the `claude --print` command to send requests to new instances, manage permissions, and aggregate results effectively. Just ensure your settings are set up correctly.