An execution layer for RL, agent tool use, and model evaluation—available today.

Until recently, models learned about the world by reading. They consumed text, predicted the next token, and produced more text. They didn't touch anything.

Today's most capable models act. They generate code, call tools, navigate browsers, and take actions on behalf of users. The training loop for agentic AI now requires execution. Agents explore action spaces and run the code they generate, and that code can be powerful, but it can also do damage.

They're running Reinforcement Learning (RL) at scale, building agentic workflows, and evaluating continuously. Today they have three imperfect choices: build their own sandbox infrastructure, buy capacity from a separate vendor, or restrict what their agents are allowed to try. Each choice compounds operational debt as research scales. Engineering hours are spent on plumbing, compute capacity is underutilized, and governance gaps increase operational and security risks.

There’s a fourth option. CoreWeave Sandboxes is an execution layer for RL, agent tool use, and model evaluation at scale, available today in public preview. Run sandboxes on your own capacity across multiple CoreWeave Kubernetes Service (CKS) clusters, or as a serverless runtime on CoreWeave-managed compute.

AI innovation needs a safe place to run

Frontier research requires letting agents explore action spaces and execute generated code. That work happens in an execution layer.

On CKS, your platform team configures sandboxes to match how you train. CoreWeave Sandboxes runs every sandbox inside its own Kubernetes pod, governed by a profile that defines network policies, resource limits, and namespace boundaries. Administrators set the boundaries once, and every sandbox a profile produces inherits them. 

On the serverless runtime, you start with code, not configuration. CoreWeave manages the cluster with strong security boundaries and runs every CPU sandbox inside a Kata Container by default, giving each sandbox its own hardware-virtualized kernel, filesystem, and network. It is one of the strongest isolation models available for these workloads, and it comes out of the box. Researchers authenticate with an existing W&B API key, install the Python client, and start running sandboxes in minutes. No clusters to provision and no infrastructure decisions to make. Your ML engineers stay focused on running sandboxes in a secure, enterprise-grade environment.

One failure stays in one sandbox. Memory spikes, infinite loops, and OOM errors are confined to the pod that produced them. Researchers scale rollouts to the thousands without one rollout's failure reaching the next.

“CoreWeave Sandboxes solves a real gap in our AI research stack: secure, isolated code execution at scale directly in our existing compute. Our reinforcement learning workflows spin up thousands of sandboxes in parallel per training step, each with its own container image and resource boundaries. Adoption is frictionless—researchers run sandboxes within minutes of a

pip install cwsandbox,

no infrastructure knowledge required.”

Brian Belgodere, Senior Technical Staff Member: AI/ML Systems, IBM Research

Iterate while you run

When agent rollouts fail at scale, the information needed to find the failure mode usually lives across disconnected systems: sandbox lifecycle events in one, training metrics in another, LLM traces in a third. When 37 of 1,000 evals fail, finding which 37 means correlating timestamps across systems that don't share context. Debugging costs compound with run size.

Every sandbox is part of the training run. Sandbox lifecycle events land in the same W&B run timeline as your metrics, indexed alongside the rollouts they belong to. W&B Weave traces connect to per-sandbox execution, so every model call, tool call, and return value is correlated with the sandbox that produced it. The agent's code, inputs, and outputs are captured per sandbox and stored alongside the run, queryable later.

You don't switch systems or reconstruct context from infrastructure logs. You stay in the run.

Get more from the compute you already have

Capacity is tight, and RL training pushes sandbox counts into the thousands per step. When routing to a separate sandbox vendor, sandboxes add inter-cluster latency, a second billing model, and another capacity ceiling on top of the one you're already pushing against.

On CKS, CoreWeave Sandboxes schedule across the clusters you already operate, picking up capacity wherever it opens up, including idle CPUs on your GPU nodes when you run them alongside SUNK. Profiles define the compute footprint each sandbox can access: instance type, CPU and memory, storage. 

Managing separate clusters and scheduling sandboxes across different node types lacked a unified solution, costing us time and resources. CoreWeave Sandboxes eliminated that issue. We now run hundreds of concurrent sandboxes on CPU nodes and alongside Slurm training jobs on GPU nodes, all through a single setup. The Python SDK let our researchers get started immediately, and the CoreWeave team worked closely with us to adapt the open-source SDK to fit seamlessly into our codebase.

Roman Soletskyi, AI Scientist, Mistral

Start running sandboxes

CoreWeave Sandboxes is available today through two consumption paths, both built on the same Python client and the same CoreWeave-managed control plane.

1from cwsandbox import Sandbox
2
3with Sandbox.run() as sandbox:
4    result = sandbox.exec(["echo", "Hello Sandboxes!"]).result()
5    print(result.stdout)
6

Because serverless sandboxes run through Weights & Biases, the platform features come built in. Every sandbox is pre-authenticated to your Weights & Biases identity, so calls to W&B Models and W&B Weave from inside the sandbox already know who you are. No API keys to pass, no credentials to mount. Secrets your agents need are injected securely from your W&B Team secrets store. Sandbox lifecycle events land in your W&B run timeline, and W&B Weave traces connect every model call and tool call back to the sandbox that produced it. Usage is tracked per org and included in your W&B plan, so there's no separate billing relationship to manage.

1from wandb.sandbox import Sandbox
2
3with Sandbox.run() as sandbox:
4    result = sandbox.exec(["echo", "Hello Sandboxes!"]).result()
5    print(result.stdout)
6

The right execution layer for AI development

The next generation of models won't be trained once. They'll be improved continuously, through RL, agent tool use, and large-scale evaluation. The infrastructure underneath has to keep up.

CoreWeave Sandboxes is the execution layer for that work, on your own clusters or as a serverless runtime.

Get started

Existing CoreWeave customers can enable sandboxes directly in the Cloud Console.

Teams looking for self-service access can get started through Weights & Biases: Docs

pip install wandb[sandbox]