CoreWeave has launched Forge, a unified development layer that connects training, inference, observability, evaluation and agent development in one continuous AI improvement loop.

Forge combines Weights and Biases Models for experiment tracking, OpenPipe’s post-training expertise, the marimo open-source notebook project and CoreWeave’s infrastructure. New capabilities include the ARIA coding agent, now generally available, which analyses experiment data and proposes the next experiments to run.

Also new is Agent Lens, an intelligent observability tool for production agents that detects 20% more critical failures and fixes issues at one-tenth the cost, based on a benchmark against a general-purpose frontier model. CoreWeave Sandboxes provide fresh, isolated environments for every agent tool call, reinforcement learning run or evaluation.

Serverless reinforcement learning trains 1.4 times faster at 40% lower cost than a self-managed setup, the company said.

Why it matters: Forge lets what a business learns from running AI feed directly back into improving its models and agents.