Summary
Sam Witteveen explains how small decision models such as Jev, and the open decision models, can improve an AI agent harness from the inside. Most model calls in an agent loop are not writing text but making decisions, such as which tool to use, which skill to load, whether a command is safe, or which of 20 search results answers the question. Using a full LLM call for each of these adds cost and latency, which is why many agents skip verification steps.
The video describes Jev as a smart if statement: it takes a state and a set of typed questions, then returns typed answers with probabilities and no generated text. Question types include choice, score and yes-or-no probabilities, and many questions can be answered in one parallel call. Witteveen outlines six places to use it in a harness, including model routing, tool selection, safety checks and result verification, and demonstrates progressive disclosure across many skills and re-ranking of RAG results.
He compares two patterns, decision hooks alongside the LLM and Jev as the primary decider with the LLM as fallback, and notes that LangChain and Pydantic have shipped related middleware. The video also covers where these models do not fit: text generation, multi-step reasoning, very long inputs, image inputs on proprietary models, and compound questions.
📺 Source: Sam Witteveen · Published September 29, 2026
🏷️ Format: Deep Dive
