We tested NVIDIA NeMo Switchyard’s stage router for coding agents in Ramp SWE-Bench. Routed agents showed comparable performance to single-model controls while substantially reducing costs and runtime.

Try Switchyard in Ramp Router →

Very cool piece of tech, been great working with the NVIDIA team on this!

So nice of Ramp to collaborate with lesser-known AI startups ❤️

@JensenHuang woahhhh this is some crazy stuff?!? You seeing this 👀

self driving spend management , dream team

Shah of benchmark @hshaiyon 🐐

The interesting part is that routing starts turning model choice into an optimization problem, not a product decision.If agents can dynamically pick the cheapest model that still clears the task, the real moat may shift toward orchestration: knowing when to spend frontier-model compute and when not to.

Similar approach in Campfire's agent race mode, but inverted: instead of a pre-defined routing rule, you run multiple coding backends concurrently and pick the best result. Campfire supports 7 backends: Claude Code, Codex, Goose, Aider, OpenHands, OpenClaw, and OpenCode. The cost-cascade variant stops remaining backends the moment the cheapest one completes successfully. If the low-cost backend solves it, the expensive ones never finish. What you get over a fixed router: no need to pre-classify which model handles which task type. The race gives you the answer empirically on each individual problem.

Comparable performance at lower cost is exactly the tradeoff teams want.

@grok what is Nvidia Nemo switchyard?

this is the right move. single-model agents waste tokens on easy stuff and still fail on the hard stuff. curious how you're handling the handoff though — does the router see the full context or just the task description? we kept hitting cases where the cheap model needed expensive-model context to even know it was the wrong model for the job.

Comparable performance with lower cost is a huge win.