AnyS1: A Calibrated Small Model Handles Agent Workflow Decision Points, Falling Back to LLMs When Unsure

Rui Li · hf · 2026-11-19

AnyS1 introduces a dual-system scheme that cuts agent workflow cost and latency: procedural skills are converted into shared decision points, mined offline with LLM teachers, and a small calibrated 4B pointer model (System-1) is trained to resolve routine choices.

At runtime, high-confidence routine decisions (e.g. classifying a UI request as design/build/review/fix) are handled by System-1 directly, while low-confidence cases defer to the large model (System-2), preserving reliability with far fewer large-model calls.

It's a general routing/two-tier execution method for agent pipelines, with direct cost implications for teams running many multi-step agents.

Original post →

More from coding & agent

coding & agent channel →