In this article, we will look at how that learning actually happens, starting with why instruction-following alone falls short, then walking through the two main methods for teaching preferences (RLHF and DPO).
Production workflows expose a limit in the verifier framing: “did the code pass?” can be deterministic; “should this customer record change?” usually cannot. Teams need to separate machine-checkable assertions from policy or judgment gates, then preserve both signals in the action record. Otherwise one green score starts doing several jobs it was never hired for.
A model trained on past agent decisions isn't really avoiding adjudication, it's just reusing old ones, inconsistencies and all.
Production workflows expose a limit in the verifier framing: “did the code pass?” can be deterministic; “should this customer record change?” usually cannot. Teams need to separate machine-checkable assertions from policy or judgment gates, then preserve both signals in the action record. Otherwise one green score starts doing several jobs it was never hired for.