what do we think about agent-native programming languages? codex/claude have some intriguing ideas but the risk is of course the models just get better and so it’s unnecessary steelman is LangSec philosophy applied to agent harnesses