Sandbox

A sandbox is an isolated execution environment for running code (often model-generated or user-supplied) with controlled access to the host system: filesystem/workspace, network, secrets, and compute.

Details

In agent systems, a sandbox is where tools run (shell commands, code execution tools, file edits) so the agent can operate on a filesystem (often as a filesystem agent) without giving it full host privileges.

Typical controls: per-sandbox filesystem (ephemeral or mounted workspace), read/write allowlists, network egress rules, CPU/memory/time limits, process isolation, and audit logs. These controls are a key mitigation for unauthorized code execution and help limit the blast radius of tool misuse and persistence attacks.

Sandbox services

A sandbox service is a hosted service (usually exposed via an API) that provisions and manages sandboxes on demand and provides primitives to run commands/code, read/write files, and stream logs. Sandbox services typically add lifecycle management (create/destroy, pooling), prebuilt images/templates, workspace mounting/sync, secrets injection, network controls, timeouts/quotas, snapshotting, and observability hooks. They are commonly used to run remote filesystem agents that need a longer-lived project directory and controlled tool execution.

Examples

• Docker-based local sandboxes
• Firecracker microVMs
• WASM runtimes
• In-memory language interpreters
• E2B (sandbox service)
• Vercel Sandbox (sandbox service)

Synonyms

execution sandbox, agent sandbox, sandbox-as-a-service, remote sandbox