Builder replaces value model with an LLM call, routes estimator outputs to the trainer

cephaloform · x · 2026-08-18

A first-hand engineering update on an agent training system. The previous reward setup relied on a naive sliding window plus guideline-based turn judgments with EMA-based advantages, which accidentally clipped tasks. Now the model itself decides window size using tools, fixing that issue.

Next up: routing the value estimator's outputs to the trainer so it co-evolves with the agent better than RAG with timestamped memories. The author jokes that they once quipped value models would be replaced by a single LLM call — and now they're actually doing it.

Original post →

More from coding & agent

coding & agent channel →