CoreWeave adds sandboxes for AI

CoreWeave on May 14 introduced Sandboxes, a new managed execution layer for running reinforcement learning, agent tool use and model-evaluation workloads in isolated environments. The company said the product is available in two forms: on-cluster inside CoreWeave Kubernetes Service and serverless through Weights & Biases. (coreweave.com) CoreWeave and Weights & Biases describe the service as a way to run agents and model-generated code in environments that start clean, stay isolated and can be discarded after use. That matters because the product is not just another GPU offering. (investors.coreweave.com) CoreWeave is packaging execution control, isolation and experiment tooling together, using its Kubernetes service for platform teams and Weights & Biases as the access layer for researchers who do not want to manage cluster infrastructure. (docs.coreweave.com) The company’s documentation says Sandboxes are in public preview. ### What exactly did CoreWeave launch? CoreWeave said Sandboxes is an “execution layer” for secure, isolated environments used in AI research and platform work. The launch materials name three target workloads: reinforcement learning, AI agent tool use and model evaluation. (coreweave.com) CoreWeave’s product page says the service can run on existing clusters or as a serverless runtime managed through Weights & Biases. Its documentation describes Sandboxes as hosted on CoreWeave Kubernetes Service capacity and reachable through a Python client. ### How are users supposed to access it? The two access models are the key part of the rollout. CoreWeave said platform teams can deploy Sandboxes on-cluster through CoreWeave Kubernetes Service, while researchers and applied AI teams can use a serverless version through Weights & Biases. (investors.coreweave.com) Weights & Biases documentation says Serverless Sandboxes is in public preview and gives users on-demand isolated environments they can create and discard with Python. (coreweave.com) Investing.com’s summary of the launch, citing the company release, said users can authenticate with a Weights & Biases API key without provisioning a cluster. ### What is bundled beyond raw compute? (coreweave.com) CoreWeave’s documentation says the platform provides a unified API for provisioning, managing and monitoring sandboxes, with a control plane and a runner deployed to a customer’s Kubernetes cluster. Investing.com’s report on the launch said the product includes session management, storage integration and monitoring capabilities. Weights & Biases ties the offering to its broader experiment and evaluation stack. (investors.coreweave.com) Its site says CoreWeave Sandboxes are for running agents and model-generated code safely at scale, while W&B Weave is positioned around tracing, evaluation and monitoring. Another W&B page says sandboxes help start each run from a consistent state so results are easier to trust, compare and reproduce. (docs.wandb.ai) ### Why does the Weights & Biases link matter? CoreWeave completed its acquisition of Weights & Biases in 2025, and both companies have since described joint product work as a combination of infrastructure and developer tooling. The new sandbox product is one of the clearest examples yet of that integration: compute isolation on one side, experiment access and observability on the other. (docs.coreweave.com) That setup is useful for teams testing agent behavior. Because the environments are isolated and disposable, they are suited to running tool-using agents, evaluation jobs and model-authored code without exposing shared systems to leftover state or dependency conflicts, according to CoreWeave and Weights & Biases materials. (wandb.ai) ### Who is the product aimed at first? IBM Research’s Brian Belgodere said in launch coverage that CoreWeave Sandboxes filled a gap for “secure, isolated code execution at scale” inside the group’s existing compute environment. The same coverage quoted Mistral AI scientist Roman Soletskyi as saying the company could run hundreds of sandboxes on CPU nodes and operate them alongside training jobs on GPU nodes. (investing.com) CoreWeave’s next step is public-preview uptake. Its docs say access is available through a CoreWeave account team or support, while Weights & Biases has published serverless documentation for users authenticating through its platform. (docs.coreweave.com) (hk.investing.com) (coreweave.com)