Sandbox Runtime

The sandbox runtime is the execution layer for AI agents. It provisions isolated environments, manages session execution, and streams events so workflows can run.

There are two paths:

• Hosted sandbox infrastructure: apps use the Sandbox SDK/API against Tangle-managed orchestration when they have an API key and product access.
• Protocol-backed sandbox services: operators run sandbox blueprints as service instances, and apps verify chain/service state plus live operator health before routing users in.

Isolation, policy, and streaming control

• Isolation and containment for untrusted or semi-trusted workloads.
• Policy enforcement for tools, data access, and budgets.
• Execution control with queueing, timeouts, and retry-aware failures.
• Streaming observability with real-time events, file updates, and execution metadata.
• Capacity management with host health, pooling, and optional autoscaling.

Where to start by role

• Workbench users who need isolated, repeatable execution. Start in the workbench.
• Operators who host runtimes and earn for reliable execution. Operator onboarding.
• Platform teams who manage execution reliability and safety. Review architecture and orchestration.
• Security teams: Start with sandboxing and safety.

Hosted runtime access is gated by product/API-key access. Protocol-backed runtime access depends on a registered blueprint service instance and reachable operator endpoint.