Cognition AI and CoreWeave Outline Infrastructure Needs for Continuous Agent Learning

Artificial intelligence agents are altering enterprise compute demands by blending live inference with perpetual model training, according to executives from Cognition AI Inc. and specialized cloud provider CoreWeave Inc. Speaking at the Fully Connected event hosted by theCUBE Research and reported by SiliconANGLE (https://siliconangle.com/2026/10/01/cognition-scales-ai-agent-infrastructure-coreweave-fullyconnected/), the leaders outlined the infrastructure requirements necessary to sustain always-on development agents.

Cognition, which develops the Devin autonomous coding assistant, is expanding the tool beyond initial code generation into long-term software maintenance, including planning, reviewing code, and resolving live production incidents. Silas Alberti, head of research and founding team member at Cognition, stated that supporting these tasks requires systems capable of continuous learning rather than isolated batch updates.

To manage ongoing reinforcement learning workloads, Cognition distributes training across data centers spanning multiple countries and continents. Alberti noted that maintaining cluster stability across thousands of graphics processing units is crucial, because a failure in a single replica can halt an entire distributed training run. Meeting these operational demands has made achieving 99.99% infrastructure uptime a primary requirement.

Cognition is also using early access to Nvidia Corp.’s Vera Rubin platform to evaluate kernel dynamics and adapt model architectures, aiming to improve price-performance ratios as hardware generations advance.

Alongside the discussion, CoreWeave introduced CoreWeave Forge, an infrastructure platform designed to bridge inference, telemetry observation, data curation, model refinement, and evaluation. Chen Goldberg, executive vice president of product and engineering at CoreWeave, said the system is tailored for end-to-end agentic feedback loops, starting from serverless inference up through complex architectures.

Forge includes Agent Lens for tracing agent operations, model distillation tools, reinforcement learning capabilities, and an RL Rollouts feature that can hot-load updated model checkpoints directly into active environments without requiring full system redeployments.