Claude code already fans out and sandboxes context by calling sub agents so I'm not sure this approach brings much benefit there. A complex search strategy only makes sense if the search is slow and compute intensive.
i find claude subagents are incredibly slow and the vast majority of that is inference time between tool calls
i find claude subagents are incredibly slow and the vast majority of that is inference time between tool calls