CoreWeave Forge aims to speed up the AI improvement loop

CoreWeave is pitching Forge as a way to make AI systems less of a one-off build and more of a living process. The idea is simple enough: run an agent, watch what it does, fix what breaks, test again, then keep going. Susanne Seitinger, CoreWeave’s vice president of product marketing, says the point is to help teams move from their first agent to their best one without having everyone reinvent the same lessons.

That matters because enterprise AI is no longer just about launching something flashy and calling it done. Once agents are in production, the real work starts. CoreWeave says Forge is built to give business units and machine learning teams a shared workflow, so the people shipping the system and the people improving it can actually speak the same language. A lot of AI tooling still feels like it was assembled in different rooms by different people. Forge is CoreWeave’s answer to that mess.

The platform breaks the loop into five stages: run, observe, curate, improve, evaluate and repeat. CoreWeave is pairing that with a set of specific tools. Agent Lens is for inspecting agent behavior. Registry tracks checkpoints and agent configurations. RL Rollouts and model distillation are aimed at improvement. And ARIA, CoreWeave’s AI Research and Iteration Agent, is now generally available for finding patterns in the work.

Seitinger framed speed of learning as the real metric. The faster teams learn, the faster they ship value, she said. That logic fits a market where post-training and inference are part of the ongoing job, not a cleanup phase at the end. She pointed to Cognition AI Inc. running production workloads on CoreWeave’s first Nvidia Corp. Vera Rubin NVL72 racks as evidence that this kind of continuous tuning is already happening.

CoreWeave is also pushing a partner network built on tested, co-engineered integrations. The pitch there is less glamorous but probably just as important: give customers more recipes, more playbooks, and fewer dead ends. In other words, make the boring parts of enterprise AI less painful. That’s where a lot of the value hides.