It makes sense to me that agents with known tool sets should be able to safely create ad-hoc "tool recipes" that are not arbitrary code execution but are kind of arbitrarily composable from allowlisted functions. After reading thru this thats what this appears to be.
I guess the addition of the json call graph makes this easier to wrap in a durable execution primitive for easy restarts.