Skip to content

Parallel and cloud environments

Harbor abstracts the environment behind a provider. The same task moves from your laptop to the cloud by changing --env.

终端窗口
# local Docker (default)
harbor run -d terminal-bench@2.0 -a claude-code -m anthropic/claude-opus-5-5 -n 4
# Daytona
harbor run -d terminal-bench@2.0 -a claude-code -m anthropic/claude-opus-5-5 -n 100 -e daytona

Daytona, Modal, LangSmith, Blaxel, Novita Sandbox, Tensorlake, Runta, Sail.

  • One task exposes one failure mode; a thousand tasks expose a distribution
  • RL needs rollouts, and rollouts need throughput
  • Re-running across providers is how you test whether the task itself is sound