Why Capital One built its multi-agent AI platform around open-weight models

At Capital One, Kel Vanee, a machine learning engineering MVP, explained that the bank's scalable multi-agent AI architecture relies on customized open-weight models rather than off-the-shelf foundation models. Capital One's investment in data transformation and cloud adoption laid the groundwork for this approach. Customizing open-weight models using proprietary data has proven advantageous, as it provides a unique edge over general frontier models.

This customization allows Capital One to train the models to be experts in Capital One's policies and nomenclature, resulting in general performance improvements. The multi-agentic AI workflow, named MACAW, routes interactions through specialized agents with built-in governance and guardrails. The workflow includes an understanding agent, reasoning agent, validation agent, and explaining agent.

For example, in a customer service workflow for bank fraud, this system helps several hundred customer-service agents manage complex fraud calls. The post-call summaries generated by the workflow reduce the time agents spend reconstructing interactions. Capital One also employs an agentic research system to optimize latency and cost in backend hosting infrastructure.

This system searches for the best combination of optimizations and configurations to achieve optimal performance. Vanee predicts a trend toward routing abstraction layers that validate the use of multiple models to achieve better accuracy and cost-efficiency. Additionally, he foresees a shift towards proactive, event-driven AI, where AI systems act independently when conditions warrant action.