Engine Stack

01 · SITUATION

One model doing everything is how AI budgets die. Planning needs a frontier model; classifying an email does not. The mandate: an orchestration layer that routes each step to the cheapest model that can do it correctly, without the layers collapsing into spaghetti.

02 · SYSTEM

Three tiers with typed contracts at every boundary: a planner that decomposes objectives, a dispatcher that routes steps by complexity to local or cloud models, and executors that do the work and report back in a schema the compiler checks. Async model serving with backpressure keeps local GPUs saturated without queue collapse; every contract violation is caught at the boundary, not in production.

03 · RESULT

Routine steps run on local models at a fraction of cloud cost, frontier calls are reserved for the steps that earn them, and adding a new executor is a typed interface away. The architecture is the backbone of the T3-02 Agent Orchestration offering.