LangSmith Sandboxes Enable Safe Agent Code Execution with MicroVMs

LangChainLangChain

LangChain has made LangSmith Sandboxes generally available, providing kernel-isolated microVMs for AI agents. These environments allow agents to safely execute untrusted, model-generated code, enabling complex workflows without compromising infrastructure security. This shifts agents from answering questions to autonomously performing work in production environments.

LangSmith Sandboxes are now generally available, offering secure, scalable environments for AI agent code execution. Each sandbox runs as a hardware-virtualized microVM (micro-virtual machine), kernel-isolated from services and other sandboxes, ensuring security for untrusted, model-generated code. The sandboxes integrate with the Deep Agents SDK and the LangSmith platform.
Availability
Generally Available
Isolation
Hardware-virtualized microVMs, kernel-isolated
Key Features
Snapshots, cheap copy-on-write forks, Service URLs, Sandbox CLI, Auth Proxy
Integration
Deep Agents SDK, LangSmith platform
Security
Creator-private by default, Auth Proxy for credentials
Customer Use
monday.com's Sidekick AI assistant

The need for strong isolation arises from risks like supply-chain attacks, such as the Shai-Hulud npm worm, and kernel exploits like Copy Fail (CVE-2026-31431), which can bypass container-level security. Containers, designed for vetted application code, are insufficient for stateful agent workloads that install packages, edit files, and follow long-running tasks with untrusted code.

LangSmith Sandboxes include features like snapshots and cheap copy-on-write forks for parallel branches, Service URLs for authenticated HTTP access, and a Sandbox CLI for management. monday.com uses these sandboxes for its Sidekick AI assistant, allowing it to write and run code for advanced user workflows, including data analysis and multimedia generation.

monday.com leverages LangSmith Sandboxes to enhance Sidekick AI assistant capabilities through secure code execution and data analysis.
LangSmith Sandboxes performance update — the platform now achieves a median spin-up time of 0.98 seconds. This capability allows developers to dynamically scale to thousands of concurrent sandboxes without managing underlying compute infrastructure, addressing the need for low-latency, high-concurrency environments for user-facing AI agents.
View the full update on langchain.com
LangChain
LangChain
@LangChain
X

Sandboxes are already helping teams move from agents that answer questions to agents that can do work safely. At @mondaydotcom, that means giving Sidekick a secure environment to write and run code for more advanced user workflows. https://t.co/C7j8aZlkMx

4retweets27likes
View on X

Still wondering? A few quick answers below.

LangSmith Sandboxes are secure, scalable execution environments for AI agents, now generally available. They use hardware-virtualized microVMs that are kernel-isolated from other services and sandboxes, allowing agents to safely run untrusted, model-generated code.

AI agents often generate and execute code, install dependencies, and manage state. Running this untrusted code requires strong isolation to prevent supply-chain attacks or kernel exploits from compromising infrastructure, which standard containers cannot fully guarantee.

The General Availability release adds snapshots and cheap copy-on-write forks for parallel execution, Service URLs for authenticated HTTP access, a Sandbox CLI for management, and an Auth Proxy with custom callbacks for secure credential injection.

Teams are using LangSmith Sandboxes to enable agents to perform complex work safely, moving beyond simple question-answering. For example, monday.com's Sidekick AI assistant uses them to write and run code for advanced user workflows like data analysis and multimedia generation.

Sandboxes are kernel-isolated microVMs, preventing code from escaping to the host system. They are creator-private by default, and an Auth Proxy handles outbound requests, injecting credentials at the network layer so secrets never touch the runtime environment.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Share this update