I'm not sure "wasted work is cheap" really holds up when you look at the token unit economics for Claude, especially if you're hitting the API with full context for every parallel agent. I switched from a similar parallelized setup to LangGraph precisely because the merge conflicts and redundant reasoning steps were killing my margins. How are you handling the git conflicts without a coordinator?