LangChain: LangSmith Sandboxes Now Generally Available — Its Own Computer for an AI Agent

LangChain announced the general availability (GA) of LangSmith Sandboxes on its blog on June 5, 2026, in a post by Amy Ru. These are isolated environments that give an AI agent, figuratively speaking, its own computer — a complete workspace in which the agent can operate without access to production infrastructure.

What are LangSmith Sandboxes?

LangSmith Sandboxes are isolated, hardware-virtualized environments implemented as microVMs (micro virtual machines), emphatically not as containers. Each sandbox gives the agent its own file system, a shell, a package manager, and persistent state that survives between steps.

Within such an environment, the agent can install dependencies, execute code, and maintain session continuity — all isolated from the rest of the system. This solves a common problem: agents need a place where they can experiment to do real work without endangering production systems.

Why microVMs instead of containers?

The decision to use hardware virtualization instead of container isolation is key from a security standpoint. Code generated by a language model is by nature untrusted — it may contain errors or behave unexpectedly. Hardware virtualization provides stronger protection against such code because it isolates the environment at the hardware level, not just at the operating system level.

Speed is not sacrificed in the process: LangChain highlights instant startup at the level of serverless functions. In other words, the agent gets strong isolation but without the slow boot of a full virtual machine.

What advanced capabilities do Sandboxes bring?

The feature set goes well beyond a mere isolated terminal. Snapshots allow saving the state of the environment, and forks work on the copy-on-write principle, which quickly creates parallel copies without duplicating all the data.

There are also blueprints for a pre-warmed environment, service URLs for local servers the agent starts, and credential injection via an auth proxy, so that sensitive data is not exposed directly to the agent. All of this integrates with the existing LangSmith SDK and API, so teams already using LangSmith do not have to change tools.

What use cases is this intended for?

LangChain cites several typical scenarios. Among them are coding assistants, which need a safe space to write and run code, and CI agents (continuous integration) that automate building and testing.

Also mentioned are data-analytics pipelines and parallel RL training environments (reinforcement learning), where multiple isolated instances can train or experiment simultaneously. In each of these cases, the key is the combination of isolation, speed, and preserved state, which Sandboxes offer through a single interface.

By announcing the general availability of LangSmith Sandboxes, LangChain positions this infrastructure as a standard building block for agents that must execute code safely, quickly, and with state preserved throughout an entire session. For teams building autonomous agents, it is especially important that security rests on hardware virtualization rather than lighter container isolation, since executing model-generated code is precisely one of the greater security concerns in developing agentic systems.