CoreWeave Inc. (NASDAQ: CRWV) announced the launch of CoreWeave Sandboxes, a service providing secure, isolated environments for AI researchers to conduct reinforcement learning, agent tool use, and model evaluation.

The service offers two access models: on-cluster deployment for platform teams using CoreWeave Kubernetes Service and serverless access through Weights & Biases for researchers who prefer not to manage infrastructure directly.

CoreWeave Sandboxes runs within a customer's existing CoreWeave cluster and includes a Python SDK for creating isolated environments capable of handling multiple concurrent jobs. The service features session management, storage integration, and monitoring tools.

For users without existing CoreWeave infrastructure, the service is accessible as a serverless runtime through Weights & Biases. Users can authenticate with their W&B API key and begin running sandboxes without cluster provisioning requirements.

IBM Research's Brian Belgodere said the service addresses "secure, isolated code execution at scale directly in our existing compute" for reinforcement learning workflows that require thousands of parallel sandboxes per training step.

Mistral AI scientist Roman Soletskyi noted that the service eliminated the need to manage separate clusters and scheduling systems, allowing the company to run hundreds of concurrent sandboxes on both CPU and GPU nodes through a single setup.

Chen Goldberg, executive vice president of product and engineering at CoreWeave, stated that the service "closes the execution gap in reinforcement learning and agent workflows without requiring teams to build custom execution systems."

The service is available through CoreWeave's Cloud Console and Python SDK, according to the company's press release.