How do you run AI agent evals in parallel on cloud GPUs?

Run each agent eval episode in its own sandbox with a per-run cap and a batch cap, grade answers separately, and size the batch for a useful error bar.