Always-on AI agents turn infrastructure into a continuous learning loop

AI agent infrastructure is evolving to facilitate continuous learning and feedback loops throughout the software development process. Cognition AI Inc.'s Devin AI agent now assists with planning, coding, reviewing, and responding to production issues, requiring robust AI infrastructure capable of scaling and maintaining continuous learning.

Silas Alberti, head of research and founding team at Cognition, emphasized that their systems are always on, continuously training and improving their models by seeking new data and reward signals. AI agent infrastructure must support both inference and training, particularly in reinforcement learning workloads where models generate responses and refine themselves based on the outcomes.

The continuous training and inference are increasingly intertwined, especially in reinforcement learning. Cognition AI has distributed training across data centers in multiple countries and continents, ensuring high uptime for thousands of graphics processing units. A single GPU failure can halt an entire training run, making a 99.99% reliability metric crucial.

Access to cutting-edge hardware, such as Nvidia's Vera Rubin platform, shapes Cognition's research direction by providing early access to new technologies and enabling researchers to study system dynamics and optimize model architectures. With each hardware generation, price performance improves, allowing more compute to be utilized effectively.

CoreWeave Forge, announced during the Fully Connected event, connects various aspects of AI agent infrastructure, including inference, observation, data curation, model improvement, and evaluation. Agent Lens provides traceability for agent activity, while model distillation and reinforcement learning capabilities facilitate continuous learning. Forge's RL Rollouts service allows hot-loading of updated model checkpoints into live deployments without the need for redeployment.

The continuous loop of learning and improvement is critical for Cognition's Devin AI agent. By leveraging real-world experience, Devin becomes a more capable entity in long-running software projects, handling production issues, generating pull requests, and merging them as needed. This continuous learning loop enables Devin to act as a real production maintainer, enhancing software systems through ongoing adaptation and refinement.