Cognition and CoreWeave executives say always-on AI agents require continuous learning infrastructure

Alberti said Cognition's training runs are always on. "Our runs are always on," he said. "While we still ship releases, package them up a little bit, I think the reality is we're always training. We're always trying to find the next data and the next reward signals to improve our models."

Training and inference are increasingly intertwined, especially in reinforcement learning workloads that generate responses and use the results to improve a model, according to the discussion. Cognition has distributed training across data centers in multiple countries and across continents, making uptime across thousands of graphics processing units essential, Alberti said.

"If you run a big training run, you're not just looking at the GPUs, but you also want uptime," he said. "If just one replica goes down, the whole training run goes down. I think a very important metric is getting to this 99.99% reliability."

New hardware also shapes Cognition's research direction, Alberti said. Early access to Nvidia Corp.'s Vera Rubin platform allows its researchers to study the system and adapt model architectures while pursuing better price-performance. "With each generation, price performance just goes up, so we can do more with the same amount of compute," Alberti said. "Being early and actually being able to study the kernels and the dynamics of this new platform allows us to prioritize our research investments."

CoreWeave announced Forge during the event. The platform connects inference, observation, data curation, model improvement and evaluation. It includes Agent Lens for tracing agent activity, along with model distillation and reinforcement learning capabilities. Its RL Rollouts service can hot-load updated model checkpoints into a live deployment without redeploying the system.

Goldberg said Forge is designed for the loops used to build agents. "CoreWeave Forge is … tailored for those AI loops, those learning systems of how you build agents and connect all the different steps," she said. "You don't have to do everything at once. You can always start with serverless inference, find your model and start."

For Cognition, the value of the continuous loop lies in using real-world experience to improve subsequent versions of Devin. The goal is to make the agent more capable across long-running software projects, Alberti said.

"Agents are not just planning the code, writing the code [and] reviewing the code, but also becoming real production maintainers," he said. "If there's a production issue, Devin can jump on it, respond to it. And by the time you wake up, there's already a [pull request] and you can merge it."

TheCUBE is a paid media partner for the Fully Connected event. Neither CoreWeave, the sponsor of theCUBE's coverage, nor other sponsors have editorial control over the content on theCUBE or SiliconANGLE, according to the disclosure.