How it works

Agent / harness → calls → task model (weights). We sell the model under the agent.

1. 1

   Start from an agent that’s already struggling

   You have a harness calling a frontier LLM for a scoped job — fraud, triage, routing, review. Labels exist. Cost or reliability is the pain.

2. 2

   Measure the gap

   Run an eval against your held-out outcomes. See frontier vs base open-weight before you spend on training.

3. 3

   Train a task model

   Post-train a small open-weight model (LoRA / GRPO when rewards exist) on your proprietary outcomes — Shared, BYOC, or confidential GPU.

4. 4

   Point the agent at the new endpoint

   Deploy the adapter. Your agent/harness keeps its tools and flow; only the weak model call changes. Inference is basically free vs frontier.