Compile the Workflow Into the Weights

Current agentic frameworks (LangGraph, CrewAI, the OpenAI Agents SDK) inject full workflow logic into a frontier model's context on every turn. Expensive, wasteful and leaky, and it is only going to get worse. This paper proposes compiling the workflow directly into the weights of a small fine-tuned model instead: near-frontier quality at roughly 100x lower token cost, validated on real workflows, with proprietary procedures staying inside your model and off third-party APIs. The authors directly dismantle the reasons developers have avoided this approach. In simple words: yes, you can completely rethink how agentic products are built and deployed. And this is wild.

Sources: paper