Google open-sourced gVisor, a container runtime with a userspace application kernel. It adds an isolation boundary by handling many application system calls outside the host kernel’s direct system-call interface.
Agent Sandboxes
Also known as: AI sandboxes · Code execution sandbox · Agent execution environments
Plain English
An agent sandbox is an isolated environment where an AI agent can run commands or use software. It gives the task a workspace with controlled access to files, credentials, and the network.
Technical Overview
A sandbox can use a lightweight virtual machine or an application kernel that mediates system calls. The harness determines which files and tools are available and supplies time, memory, and network limits. For agents that need to call a service, a proxy can apply scoped credentials outside the sandbox and keep those credentials out of the process environment.
Timeline (4)
2018
Release- Release
AWS open-sourced Firecracker, a KVM-based monitor for microVMs. Its minimal device model reduces the startup and resource overhead of providing a separate guest environment.
2026
ReleaseCloudflare made Sandboxes generally available. Its SDK gives applications control of an isolated environment’s lifecycle, command execution, files, and persistent interpreter sessions.
- Release
E2B launched Workload Identity, which attaches a signed, audience-scoped identity to matching outbound requests. Services can authorize individual sandbox executions without a long-lived credential inside the sandbox.
See also (6)
Open full graphAgent Sandboxes