Ask the Experts: How NeMo Switchyard Helps Agents Select Models | Nemotron Labs

No single model excels at every task. NeMo Switchyard is an open source model routing library for AI agents that solves this by automatically routing each agent query to the best available model — selecting from any combination of closed and open models, cloud and local models, based on what each step needs, with no training or configuration required. This session goes inside the library with the product manager and engineers who built it. They'll walk through the design decisions behind Switchyard's built-in routing algorithms — how they evaluate each agent step, carry state across multi-turn sessions, and get smarter as your agents run. You'll also see how teams extend Switchyard with their own routing algorithms, adjusting routing strategies to match their specific accuracy, cost, and latency priorities. What you'll learn: How NeMo Switchyard routes each agent step automatically — and how it differs from request-level model routers How the built-in routing algorithms work: tuning-free options that start routing immediately, and a tunable option that improves with real workload signals How to run Switchyard as a standalone server, Python library, or embedded natively in an agent harness How to bring your own routing algorithms and adjust routing strategies for your specific use case How to get started and contribute on GitHub Questions about routing algorithm design or building on top of Switchyard? The engineers who built it will take them live.