Nushell in three spoonfuls: when does a structured shell actually help an agent?

I tested whether selectively routing coding-agent tasks through Nushell helps: it improved accuracy on some tasks, increased latency and output tokens, and made no accuracy difference in a real aggregate case.