Switch: an open harness that lets decision models act in loops by picking from generated next steps
publicado (OK do Fabricio 30/10 BRT).
Onur Solmaz (ex-Hugging Face) introduced Switch on 29 October 2026 (15:08 BRT), one of the most-liked decision-model posts of the week (about 1.2K likes). Switch is an MIT-licensed TypeScript framework for “decision-model harnesses”: since a decision model cannot write tool calls, the program generates a short list of possible next steps from the current state and the model picks one, step after step, until it can answer.
How it is organised. You implement a “domain”: how to represent the state and which questions to derive from it. The first domains are chess and a rubric judge for evals, where the model can page through a long document set instead of having every page dumped into context. It is not tied to Jev; any model reachable through the pi-ai package works.
What the author says about Jev. He reports that Jev is weak at chess and, in his domain, not good enough yet to replace GPT 6.1 Sol as an eval judge. His bet is to collect data from a stronger model and fine-tune an open decision model inside the harness.
Why it matters here. Switch is a harness, not a model, and it trains no weights, so it is news, not a catalog entry. It formalises the “agent loop over finite choices” pattern many teams are already coding by hand. The repo was created on 25 Oct and had about 230 stars on 30 Oct.