Why this exists
Most agent frameworks pick a side: fully local (slow, inflexible) or fully hosted (expensive, dependent). Harness Core treats the provider as a deployment decision per task, not a framework-level axiom.
What it does
Classifies incoming tasks along an internally-defined topology (n3), routes each task to whichever provider fits — local Ollama for bulk or sensitive work, hosted Anthropic for tasks that need strong reasoning. A SQLite-backed task queue keeps the whole system crash-safe.
What I learned
Classification is where this pattern lives or dies. Getting the topology right mattered more than getting the routing right — cheap/slow providers can still win when the task is correctly framed.