Cognition AI is evolving its Devin agent to support continuous learning throughout the software development lifecycle. The system now handles planning, coding, reviewing, and resolving production issues. To maintain performance, the company distributes training across data centers globally, requiring 99.99% uptime for thousands of graphics processing units.
The company is leveraging early access to Nvidia's Vera Rubin platform to optimize model architectures and improve price-performance. By studying the dynamics of new hardware, researchers can prioritize investments and increase output with the same amount of compute resources.
CoreWeave introduced its Forge platform, designed to connect inference, data curation, and model improvement. The service includes Agent Lens for activity tracing and RL Rollouts, which allows hot-loading updated model checkpoints into live deployments without system redeployment.
The collaboration between Cognition AI and CoreWeave focuses on creating a continuous feedback loop. By using real-world experience to improve Devin, the agent acts as a production maintainer capable of addressing issues and generating pull requests autonomously.
