A novel local-model branch of the fast-decision-model trend broke out on HN: Jeff is an independent Apache-2.0 family of open 0.8B/2B “System 1” classifiers that returns calibrated probabilities for tool choice, safety/triage and other bounded decisions, without generating text. Its author reports 22–60 ms GPU/Apple-silicon decisions and 78.7–81.7% across five classification-heavy benchmarks, but it is not a planner and remains materially weaker on reasoning-heavy tasks; benchmark comparisons are author-reported and Jev figures use a different sample. The project’s HN post drew 573 points/224 comments. Sources: https://github.com/firelex/jeff ; https://news.ycombinator.com/item?id=49883844
GitHub Agentic Workflows v0.90.3 adds built-in persistent work queues and typed ledgers for stateful workflow coordination, task-level Copilot model routing, GPT-6.1 Sol support, and a run-wide tool-call budget; related changes tighten safe-output handling and include a TLA+ security model with fail-closed credential cleanup. This pushes agentic CI toward auditable, bounded multi-step work rather than one-shot prompts; these are project-reported capabilities, not independent effectiveness results. Source: https://github.github.io/gh-aw/blog/2026-10-05-weekly-update/