Introducing LangSmith Sandboxes: Secure Code Execution for Agents

LangSmith Sandboxes are secure, scalable environments designed for safely running untrusted code generated by AI agents. They address the risks of letting an LLM execute arbitrary code by providing ephemeral, locked-down environments with controlled resource consumption. Integrated with the LangSmith SDK, developers can spin up a sandbox for tasks like data analysis or automated coding. Key features include custom runtime configurations, long-running sessions with persistent state, and security measures like microVM isolation and an authentication proxy to keep credentials separate from the runtime.