CoreWeave Forge Unifies W&B, OpenPipe, and Marimo Into One Platform

CoreWeave has launched Forge, a development layer that brings together the former Weights & Biases, OpenPipe, and Marimo products in one connected environment. The platform covers training, inference, notebooks, sandboxes, a model registry, and an agent observability product called Agent Lens. Existing Weights & Biases users retain their credentials and wandb.ai links continue to work. Forge is cloud-agnostic, meaning configurations carry over across wherever users run workloads, and it is available in Free and Pro tiers.

Forge includes three components at general availability: ARIA, a coding agent that automatically analyzes experiment and observability data, surfaces what drove a change, and proposes next experiments with supporting evidence; Sandboxes, isolated environments for agent tool use, reinforcement learning, and model evaluation; and Agent Lens, an observability tool for production agents. Agent Lens reads the full production corpus of agent traces rather than a sample, groups conversations that fail the same way even when wording or tool sequences differ, and links each failure group to the underlying traces. The tool's Insights tab detects trends at small failure rates such as 1% of conversations. CoreWeave states Agent Lens detects 20% more critical failures and fixes issues at one-tenth the cost, based on a benchmark against a general-purpose frontier LLM.

Corey Sanders, a CoreWeave executive, named Agent Lens as his favorite launch from the Fully Connected event and said that without observability teams are 'dead in the water'. Chen Goldberg of CoreWeave said modern AI infrastructure must integrate compute, networking, storage, cooling, security, and software as one system, and that as agents run longer and use more tools, the underlying infrastructure matters almost as much as the model itself.