CoreWeave Introduces Sandboxes for Reinforcement Learning and AI Evaluation

CoreWeave has launched a new product called CoreWeave Sandboxes, designed to provide secure and isolated environments for reinforcement learning, AI agent tool use, and model evaluation. The release was announced in a press release and is available both on CoreWeave infrastructure and as a serverless runtime through Weights & Biases.

The Sandboxes system serves as an execution layer that lets AI researchers and platform teams run code safely and at scale. It supports reinforcement learning and evaluation workflows that require persistent state and concurrent execution. Users can access the service through the CoreWeave Cloud Console or Python SDK, enabling direct integration with CoreWeave Kubernetes Service clusters.

For users without CoreWeave clusters, Sandboxes can be used through Weights & Biases by authenticating with an API key and installing the Python client. Each sandbox runs in a fully isolated virtual environment, ensuring that failures or resource spikes do not affect other workloads. Activity and metrics are captured within the same Weights & Biases interface for simplified debugging.

CoreWeave Sandboxes addresses the need for a unified execution layer that replaces custom-built or third-party systems used for AI agent workflows. It provides built-in session management, storage integration, and monitoring tools to help teams manage large-scale reinforcement learning and evaluation tasks more efficiently.