Developer Uses Small Model as Judge to Guard AI Agent Workflows

On November 2, 2026, developer Omer Bin Saleem published an article on DEV Community, with a version also appearing on his own site omerbinsaleem.com, describing how he wraps an AI agent with a small judge model. The idea is to let the small model absorb the routine decisions that surround the main work: hundreds of small yes-or-no questions a day, answered by the judge instead of by a big model, with the judge allowed to act on its own only once it clears a confidence gate.

Saleem reports that the judge answered 116,097 of those questions over 20 days, at a median of 245 ms, and that it only acts alone above the confidence gate. He gives two worked uses. The first is model routing, where the judge picks the model per message; he says the same report went from 137 s to 45 s. The second is merge request review, where findings on code the MR never touched scored between 0.19 and 0.40, while real ones scored 0.94. The piece stays at the level of the technique and its reported numbers, without naming the specific judge model, the confidence threshold, or how widely it is deployed in production.