LangSmith Sandboxes are now generally available, offering hardware-virtualized microVM environments for safely running AI agent-generated code. Unlike containers, each sandbox is kernel-isolated, protecting against supply chain attacks, kernel exploits, and untrusted model-generated code. The GA release ships snapshots with copy-on-write forks for parallel agent branches, Service URLs for authenticated HTTP access to sandbox-hosted services, a Sandbox CLI, creator-private access controls, and an Auth Proxy that injects credentials at the network layer without exposing secrets to the runtime. Upcoming features include local-to-cloud agent promotion, shared volumes for multi-agent collaboration, and full execution tracing.
Table of contents
Why do agents need sandboxes?Agent code needs strong isolationLangSmith SandboxesHow teams are using SandboxesWhat's Coming NextGet Started2 Impressions