CoreWeave launches Sandboxes for agents

CoreWeave said on May 14 that it launched Sandboxes, a new execution layer for running AI agents, reinforcement-learning workloads and model evaluation inside isolated environments. The company said the product is aimed at AI researchers and platform teams that need to execute model-generated commands, tool calls and file operations without exposing host systems or neighboring workloads. CoreWeave said Sandboxes is available on a customer’s own CoreWeave infrastructure and as a serverless runtime through Weights & Biases. The launch comes as AI developers push more agent and reinforcement-learning systems from demos into production environments that need policy controls, session management and monitoring. ### What exactly did CoreWeave release? CoreWeave described Sandboxes as an execution layer rather than a model or training service. In the company’s product page and launch post, it said the service provides secure, isolated environments for reinforcement learning, agent tool use and model evaluation at scale. The documentation says those environments are designed for workloads where models need to run commands, edit files or call tools during a session. CoreWeave said a sandbox can persist across multiple tool calls inside a single episode, which lets state carry over between steps while still keeping that session separated from other runs. ### Where can customers run it? CoreWeave said customers can deploy Sandboxes in two ways. One option is on dedicated infrastructure through CoreWeave Kubernetes Service, or CKS, where teams run isolated sandbox environments on their own compute across one or more clusters. Weights & Biases is the named partner in the second option. CoreWeave said Sandboxes can also run as a fully managed serverless runtime on CoreWeave-managed compute through Weights & Biases, giving customers a hosted path instead of managing cluster capacity directly. ### What problem is the product supposed to solve? CoreWeave’s documentation points to a specific operational issue: agent and reinforcement-learning systems often need to execute untrusted model-generated code. The company said that code can modify filesystems, reach networks or create non-deterministic results if it runs without isolation. The docs say Sandboxes uses policy-controlled environments for those workloads. CoreWeave also published configuration options for ingress and egress policies, namespace strategy, runner placement and profile binding, indicating that network and runtime controls are part of the product’s setup. ### What ships with the launch? CoreWeave said Sandboxes is available through its Cloud Console and a Python SDK. The company said the launch version includes tools for creating and managing isolated environments, handling multiple jobs at the same time, and integrating storage and monitoring into those runs. The public-preview documentation also lists administrator prerequisites. CoreWeave said customers using the CKS version need a running CKS cluster and specific identity and access management permissions, including a SANDBOX_ADMIN action to create profiles and runners. ### Why does Weights & Biases matter here? Weights & Biases gives CoreWeave a distribution path into teams already running experiment tracking and evaluation workflows. CoreWeave said the serverless runtime is available through W&B, tying the new execution layer to an existing platform used by AI researchers and engineers. The pairing also matches the product’s stated use cases. CoreWeave’s launch materials repeatedly name reinforcement learning rollouts, agent harnesses and evaluation benchmarks, which are the kinds of workloads that often sit alongside experiment management and observability tools. ### When can customers use it? May 14 is the launch date CoreWeave gave for Sandboxes, and the company’s documentation says the product is in public preview. CoreWeave said customers can request access through their account team, through CoreWeave Support or by email, while setup instructions are already posted in the Sandboxes documentation and get-started guides.