Always-on AI agents turn infrastructure into a continuous learning loop
SiliconANGLE Victoria Gayton ● Covered by 5 sources
AI agents are now running in a loop: do work, learn from it, do it again. That’s pushing infrastructure toward nonstop training, not one-off model updates.
Based on reporting by SiliconANGLE, Victoria Gayton — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AI agent infrastructure is starting to look less like a deployment stack and more like a feedback machine. Cognition AI says Devin now spans the software development lifecycle, from planning and coding to review and production response. That shift sounds neat until you hear the operational bill it creates: always-on runs, constant learning signals and a much higher bar for reliability.
Silas Alberti, who heads research and sits on Cognition’s founding team, said the company is no longer thinking in neat release cycles. The work is always moving between inference, feedback and training. In other words, the model is not just being used; it is being shaped by the work it does. For Cognition, that means chasing the next useful data and the next reward signal while the system is already live.
That live-loop approach gets serious fast when training is distributed across data centers in multiple countries and across continents. Alberti said uptime across thousands of graphics processing units matters because one replica failing can take down the whole run. He pointed to 99.99% reliability as an important target. And he tied research priorities to hardware access too, saying early access to Nvidia’s Vera Rubin platform lets Cognition study the system and adjust model architectures while aiming for better price-performance.
CoreWeave is trying to sell the infrastructure that matches this way of building. Its new Forge platform links inference, observation, data curation, model improvement and evaluation. It includes Agent Lens for tracing agent activity, plus model distillation and reinforcement learning tools. Forge’s RL Rollouts service can hot-load updated model checkpoints into a live deployment without redeploying the system, which is about as close as cloud infrastructure gets to admitting the software is never really done.
For Cognition, that matters because Devin is being pushed beyond code generation. Alberti described agents that can become production maintainers, jump on issues, and leave behind a pull request by morning. That is the real bet here: not just faster software, but software that keeps teaching itself how to stay in the loop.
My take — AI-written commentary, not fact-checked reporting
The industry keeps pretending agentic AI is a product category, when it’s really an operations problem wearing a shiny badge. Once the system is always learning, reliability stops being a nice-to-have and becomes the whole job. The funny part is that the grand AI future still depends on boring things like uptime, checkpoints and not falling over at 3 a.m.
Read more about this at: SiliconANGLE
