CoreWeave launches platform letting AI agents learn and improve autonomously in production

CoreWeave introduces autonomous improvement capabilities for AI agents

CoreWeave introduces autonomous improvement capabilities for AI agents Artificial intelligence cloud operator CoreWeave Inc. today announced the launch of a new offering that gives enterprise outfits the ability to deploy AI agents that can learn and improve themselves autonomously using real-world data. The current lifecycle for AI agents operates through a slow, iterative mechanic of running evaluations and fine-tuning based on reviewing metrics. This is because generative AI large language models, the "brains" that underpin agents, can behave differently between testing and real user scenarios. After watching this cycle take place so many times within its own infrastructure, CoreWeave decided to short-circuit the process by eliminating the bottleneck and allowing enterprises to launch agents that can learn and adapt in the field. "Most enterprises are stuck in a cycle of building and testing agents before they ever reach real users, and that cycle is becoming too slow and too expensive to sustain," said Nick Patience, vice president and practice lead for AI platforms at the Futurum Group. The new platform provides serverless reinforcement learning, a mechanism by which LLMs post-trained and fine-tuned for reliability. Powering the company's new offering is an engine that scales training for multi-turn agentic tasks without requiring enterprise companies to roll their own infrastructure. CoreWeave said it can reduce costs by over 40% and accelerates training by about 1.4 times with no loss in quality. Training and inference always run on separate instances, so iteration cycles do not compete with one another. The result is that what took hours of training can now be handled in seconds and updates in a blink of an eye. The company has already built vast scale AI inference and training cloud infrastructure to support model and agent deployment. Using CoreWeave Inference, users will be able to monitor the ongoing agentic systems and LLM fine-tuning processes to maintain reliable performance, runtime flexibility and stable behavior under real-world traffic at scale even as workloads grow. The era of agentic fleets coming into its own The early large language model era brought chatbots, which acted like simple wake-and-respond conversational interfaces to answer questions, summarize large documents and give a human-like back-and-forth. The agentic AI era brought autonomous capabilities for LLMs, where chatbots gave way to "thinking" software that could take on goal-oriented tasks within enterprise systems. Agents can break down long-term goals into sub tasks and tackle them with little or no human supervision and with each generation have been built to handle ever more complex work. According to the McKinsey & Co. State of AI in 2025, about 62% of industry respondents said they were at least experimenting with AI agents. With 88% reporting the use of AI in at least one business function, compared with 78% in 2024. Beyond experimentation, LangChain Inc.'s 2026 State of Agent Engineering noted that production momentum is real. Some 57% of respondents having agents in production, with large enterprises leading adoption and the use of multiple models under the hood becoming the norm. More enterprises are finding themselves working with multiple agents at once that call upon one another to orchestrate larger tasks. This additional complexity means that agents are being customized, run long-term and operated in changing conditions where their fine-tuning on data scales up with the number of agents in the network. CoreWeave said that its platform is designed to enable this new era by giving developers the advantage at scale. Agents don't need to move slowly into production from testing - as they traditionally have been built . Instead, they can adapt, learn and fine-tune themselves in production. As business data and tools change, the fleets of agents adjust themselves to match, the gap is smaller.

CoreWeave Speeds AI Agent Deployment With Real-World Learning | PYMNTS.com

These capabilities eliminate the need to run lengthy offline evaluations of AI agents before releasing them to real users for inference, the company said in a Thursday (May 28) press release. CoreWeave's unified agentic capabilities bring together reinforcement learning, production inference, agent observability and autonomous improvement in one closed loop, according to the release. The solution includes the company's Serverless RL, which enables the post-training of AI models for multiturn agentic tasks without provisioning or managing infrastructure; CoreWeave Inference, which is designed to operate as a controllable, continuously running workload; W&B Weave, which serves as the observability layer for the continuous loop between production behavior and agent improvement; and W&B Skills and MCP Server, which turn general-purpose coding agents into AI researchers and agent builders, per the release. The unified agent capabilities are meant to replace the way teams currently build AI because the process results in development cycles that can't keep up with the pace of AI or the shipping of agents that later fail in production, Chen Goldberg, executive vice president of product and engineering at CoreWeave, said in the release. "Enterprises that put agents in production first and let them continuously improve from real-world experience aren't just building more reliable AI, they're accelerating the path to superintelligence," Goldberg said. Phil Gurbacki, vice president of product, Weights & Biases at CoreWeave, said in a Thursday blog post that this new platform closes the gap between development and production, which has been where agent projects stall. "Improve automatically, and build systems that compound," Gurbacki said. "Enterprises that successfully close the loop will deliver the most reliable agents to users." The PYMNTS Intelligence report "Agentic AI Breaks Out of the Sandbox" found that the share of companies merely considering using agentic AI dropped considerably between August and November as more companies actively deployed the technology. As of November, among firms in the United States with at least $1 billion in annual revenues, 11.7% said they were already using agentic AI tools and another 11.7% said they were piloting these tools. The figures were up from 1.7% in August, according to the report. For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.