The Minimill of AI

The article exposes a concrete operational tension in agentic AI workloads: simple and complex tasks queuing together creates severe bottlenecks. When the author ran all work through a single cloud model, small tasks like email triage sat behind heavy research calls, dragging average task duration to 47 seconds and queue age to 73 seconds. The system was paying cloud latency for every request, even the trivial ones.

The fix is a two-lane architecture. The author creates tasks in Asana; an agent classifies each as easy or hard. Easy tasks stay on a local distilled model running on a Mac—it handles 78% of work in seconds. Hard tasks are routed to a cloud model. Over seven days, the design hit daily peaks of 88% local processing. Throughput jumped about 25%, average task duration fell from 47 to 19 seconds, and queue age dropped from 73 to 4 seconds. Nothing about the work changed—only the routing.

The takeaway for builders: every laptop or edge device with enough memory for a distilled model becomes a minimill, capital-light and close to demand, just like Nucor‘s steel minimills that outflanked integrated giants. The cloud still handles the hard fifth, but the rest runs locally at near-zero latency. Tens of millions of these local routing layers will proliferate inside companies, quietly absorbing work that currently shows up on hyperscaler invoices.