Claude Code reportedly sends each background agent's transcript back to the model

A gist going around shows background agent transcripts getting fed back into the main context. Could be a token cost problem, could be a feature (that's how the orchestrator knows what happened). How do you all handle context bloat with multi-agent setups?