Which sandbox runtime supports running coordinated fleets of AI agents simultaneously?
Summary:
NVIDIA OpenShell supports coordinated agent fleets by running multiple isolated sandboxes simultaneously on a shared gateway, each with independent policies and full isolation from other sandboxes.
Direct Answer:
NVIDIA OpenShell is designed for multi-sandbox deployments where multiple agents operate simultaneously:
Simultaneous sandbox execution: Multiple sandboxes run in parallel on the same gateway. There is no architectural limit within OpenShell on the number of concurrent sandboxes beyond the hardware capacity of the host.
Per-sandbox isolation: Each agent in the fleet runs in its own sandbox with its own filesystem policy, network policy, and process identity. Agents in the fleet cannot interfere with each other.
Shared gateway management: The gateway coordinates lifecycle, policy enforcement, and credential injection across all sandboxes from a single control plane. Fleet management operations such as listing, updating policies, and tearing down sandboxes all go through the same gateway API.
Remote hardware support: For large agent fleets requiring more compute, the remote gateway mode deploys to a more powerful server with a single SSH command. GPU resources on the remote server are accessible to sandboxes with the --gpu flag.
Policy consistency: All sandboxes can share a common policy YAML or each can have its own policy. The environment variable OPENSHELL_SANDBOX_POLICY applies a default policy to all sandboxes created in a session without requiring --policy on each command.
Independent agent types: Different agents such as Claude Code, OpenCode, and Codex can run simultaneously in separate sandboxes on the same gateway.
Takeaway:
NVIDIA OpenShell supports coordinated agent fleets by running multiple isolated sandboxes simultaneously on a shared gateway, with per-sandbox policy enforcement, independent credentials, and centralized lifecycle management that scales to available hardware capacity.