CoreWeave is launching an integrated platform that combines training and inference for AI agents. With Serverless RL, production inference, observability via W&B Weave, and autonomous improvement via W&B Skills, agents are designed to continuously learn from real-world experiences.
The approach revolves around what CoreWeave calls the “superintelligence loop”: a closed feedback loop where agents learn while running in production. “The pace of AI has outrun the way teams build for it. Today’s tradeoff: dev cycles that can’t keep up, or shipping agents and discovering failure modes in production,” says Chen Goldberg, Executive Vice President of Product and Engineering at CoreWeave.
Last year, CoreWeave acquired Weights & Biases, an MLOps toolmaker specializing in experiment tracking and model management. That acquisition now forms the technical foundation for the observability layer of the new platform.
Four building blocks
The platform consists of four components. CoreWeave’s Serverless RL, previously launched as a standalone service, enables enterprises to post-train large language models for multi-turn agentic tasks without managing GPU infrastructure themselves. According to CoreWeave, this makes training approximately 1.4 times faster and up to 40 percent cheaper, without sacrificing quality.
In addition, there is CoreWeave Inference, designed as a continuously running workload with built-in monitoring for scaling behavior and system health. W&B Weave serves as an observability layer: it detects failure modes in multi-agent workflows and provides an evaluation framework that prevents regressions as systems scale.
The fourth component consists of W&B Skills and an MCP server. This combination transforms generic coding agents into autonomous AI researchers working to improve other agents. Skills gives coding agents direct access to W&B tools for experiment tracking, tracing, and monitoring. The MCP server provides the resources to retrieve data and run experiments.
All new CoreWeave capabilities are available immediately.