What is an Agent Sandbox?

Sandboxing exists because agentic AI can execute arbitrary code, browse the web, and manipulate files on a machine's behalf. Without isolation, a single bad tool call, a prompt injection, or a runaway loop could touch the host filesystem, leak credentials across tenants, or exhaust shared resources. A sandbox constrains what the agent can see and do: a scoped filesystem, network egress rules, resource limits (CPU, memory, timeouts on non-LLM operations), and no shared state with other running agents. Most production agent platforms create sandboxes lazily, spinning up a fresh container only when an agent process actually needs to run, then tearing it down or recycling it afterward. This keeps idle cost near zero but introduces a timing wrinkle: anything that must be injected into the sandbox, like OAuth tokens for a connected service, has to be synced after the container exists, not before.