LangSmith Sandboxes hit general availability this week, and if your agents write and execute code, you should care. Each agent now gets its own hardware-virtualized microVM — a separate kernel, its own filesystem, its own network boundary. Cold starts land under one second. Credentials never touch the agent runtime. The security problem that made production AI code execution uncomfortable has a production-grade answer.
Why Your Current Setup Is a Liability
If you’re running agent-generated code inside Docker containers, you’re sharing the host kernel. That’s the whole problem. When an agent writes code with a kernel exploit, the container boundary doesn’t stop it. CVE-2024-21626 proved that runc containers can be escaped with a crafted payload.
The scale of the risk is concrete: 45% of AI-generated code fails security tests, according to a 2025 Veracode report. On a Docker-only setup, a subset of those failures is not a linting error. For anyone running agentic code review, CI automation, or coding agents with filesystem access, that tradeoff became untenable.
What General Availability Added
LangSmith Sandboxes entered preview in March 2026. The GA release ships four primitives that were not available before:
Sandbox CLI
You can now build snapshots from Dockerfiles, manage sandboxes, open interactive consoles, tunnel raw TCP, and use standard tools against a sandbox exactly like any Linux box. This is the operational interface teams needed to treat sandboxes as infrastructure.
Blueprints
Define a base environment: repo state, installed dependencies, warmed caches. New sandboxes spawn from that Blueprint. When your upstream changes, rebuild the Blueprint and all subsequent sandboxes get the update.
Auth Proxy
Outbound requests from a sandbox route through a proxy that injects credentials at the network layer. The agent runtime never sees your secrets. This matters especially for teams running agents with database access or cloud provider keys.
Snapshots and Copy-on-Write Forks
Capture a running sandbox, then boot new ones from that snapshot instantly. Forks give each parallel task independent state without the cost of a full boot. For CI-style agents running many parallel subtasks, this cuts setup from 90 seconds to under 500ms. See the full GA announcement for details.
How to Plug It In
If you are using the Deep Agents SDK, the integration is minimal:
from langsmith.sandbox import SandboxClient
from deepagents import DeepAgent
agent = DeepAgent(
model="claude-sonnet-4-6",
backend=LangSmithSandbox(client=SandboxClient()),
system_prompt="You are a coding assistant with shell access."
)
For LangGraph users, the thread-scoped sandbox pattern gives each conversation thread its own isolated state. The sandbox persists for the duration of the thread. monday.com uses this pattern for its Sidekick assistant, letting it write and run code for data analysis and multimedia generation — without touching production infrastructure.
How It Compares to E2B, Daytona, and Modal
E2B remains faster on raw cold-start numbers at a median of 264ms versus LangSmith’s sub-one-second target. See a full platform comparison in the LogRocket sandbox breakdown. For pure isolation performance on standalone workloads, E2B still wins. Daytona pushes sub-90ms. Modal has the better GPU story if your agents need compute-intensive tasks.
LangSmith’s edge is the stack integration. If you’re already using LangSmith for tracing and evals, you get that observability inside the sandbox with the same API key. The Auth Proxy for credential brokering is unique in this space: none of the standalone sandbox providers broker credentials at the network layer by default.
Pricing and Availability
The free tier includes 8 LSU per month. Paid tiers bill at 0.0576 LSU per vCPU-hour, 0.01845 LSU per GiB-hour, and 0.000123 LSU per GiB-hour for storage, all billed per second. The Developer plan caps at 10 concurrent sandboxes. No GPU support yet, though a self-hosted Helm option is in progress.
LangSmith Sandboxes are available now. The free tier requires no credit card. Full documentation including the Sandbox CLI reference is live on the LangChain docs site.

