What is a Sandboxed Execution?

A sandbox typically uses OS-level isolation — a container, a virtual machine, or a restricted process namespace — to constrain filesystem access, network egress, and system calls to an explicit allowlist. The goal is containment: if the executed code is malicious, buggy, or an AI agent takes an unexpected action, the blast radius stays inside the sandbox boundary instead of reaching production data or other tenants. This matters acutely for AI agents that can write and execute their own code or shell commands, since an LLM's output is not fully predictable — sandboxing is the safety net that makes it acceptable to let an agent actually run what it generates rather than only suggest it. Well-designed sandboxes also enforce resource limits (CPU, memory, execution time) so a runaway process can't degrade the host. Multi-tenant platforms sandbox per tenant, not just per task, so that one customer's agent execution can never read another customer's files or credentials even if both run on shared infrastructure — this is a foundational control for both security and compliance (SOC 2, ISO 27001) in any product that lets AI execute code.