I don't know what brought us here, but we're here so bring it on!
An agent that answers locally with a small Qwen model first and only calls out to Fireworks when it isn't confident, cutting token spend without giving up accuracy across all 8 task types.