Beyond the Harness: How LiquidAI Is Solving the Real Frontier Problems in On-Device Agents

Most agent harness architectures look the same whether they run in the cloud or on a device, until you actually try to ship one on a phone, a car, or a laptop. That's where three problems show up that don't have good answers yet: Routing intelligently between edge and cloud, building an edge-sized model capable enough to be the agent's main"brain" (with planning, memory, and execution), and acting on interfaces with no API or MCP tool to call. Jeffrey Li, COO of Liquid AI, walks through how Liquid is tackling each of these in production, drawing on how routing decisions actually play out in real deployments. You'll also learn how Liquid trained device-native foundation models with CoreWeave, models designed from the outset for constrained, on-device environments rather than compressed down after the fact, and how they’re thinking about governance, control, and observability as agents increasingly execute across both edge and cloud.