More than just hard budget caps, we need agents to be aware of those hard budget caps, and plan accordingly. Too much agent behavior is driven by the immediate proximal goal instead of long term, strategic considerations. What we have now is agents deciding to on embark on expensive, circuitous routs to a goal, perhaps with potential but not certain in any case, benefits, and intervention depends on either an attentive user or forces stop after the money pot dries up. Prior research (#?) has Dem nsttated agent behavior to adjust behavior in response to budget considerations that appeared to enable more efficient token usage, and if that means it needs to be imposed on a custom harness (and not the self interested ai provider), so be it.