CoreWeave's Forge Platform Streamlines the Continuous AI Agent Improvement Cycle

Companies deploying AI agents at scale are discovering that optimization is not a one-time event but an ongoing necessity. Agents require constant execution, monitoring, assessment and refinement to deliver sustained value. CoreWeave Inc. has constructed an integrated platform targeting agentic AI workloads, with Forge serving as its centerpiece for aligning business and machine learning teams around shared processes.

According to Susanne Seitinger, vice president of product marketing at CoreWeave, "It's our new way of thinking about the AI loop in a way that allows you to go from first agent to your very best agent as quickly as possible. What it does is put all these disparate pieces together so different teams can talk to each other more effectively." Seitinger outlined the vision during remarks at the Fully Connected event, where she spoke with theCUBE Research's John Furrier and Dave Vellante.

A Five-Stage Cycle for Continuous Improvement

Forge operationalizes the improvement cycle through five sequential stages: run, observe, curate, improve, evaluate and repeat. Each iteration builds upon the previous one, with the velocity of learning serving as the primary success metric. "The faster you learn, the faster you ship, the faster you get value. That comes back to the entire platform and the entire cloud that we're building," Seitinger said.

The platform introduces several new tools targeting the most challenging aspects of this workflow. Agent Lens provides visibility into agent behavior and decision-making. Registry stores model checkpoints and agent configurations for easy reference and rollback. RL Rollouts and model distillation capabilities handle the refinement phase, while CoreWeave's AI Research and Iteration Agent, or ARIA, now available in general release, assists users in identifying patterns and insights across their workloads.

Model distillation—extracting knowledge from large models and compressing it into domain-specific versions—represents a critical capability for enterprises. "How do you take the knowledge in a large model and distill it down into something that's specific to your use case? That's so important for all these enterprises and organizations," Seitinger explained.

The shift in workload patterns underscores the urgency of this approach. Cognition AI Inc. is already executing production workloads on CoreWeave's initial Nvidia Corp. Vera Rubin NVL72 racks. "This idea that you're never done, that you're going to do post-training, that you're going to continue to improve is a reality," Seitinger noted. "The expertise we have is [being] able to drive benefit even later on in the loop and when you're running inference."

Building an Ecosystem of Tested Integrations

https://www.youtube.com/embed/rAhfOujHcoM?feature=oembed

Beyond the platform itself, CoreWeave is establishing a partner network featuring pre-validated, jointly engineered integrations. This ecosystem approach aims to reduce the time enterprises spend reinventing solutions. "It's really about giving folks more recipes and playbooks and use cases and solutions so they can get to value faster," Seitinger said. "I want to help on our end, with the marketing work, provide more playbooks that folks can deploy quickly so that they don't have to have all the same learnings all over again."