I think this is a reasonable take for allowing agents with static capabilities to operate faster and more efficiently. I think that interactive flows may not benefit from this as much as just having a plethora of cli tools and a sandbox.
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.
I think this is a reasonable take for allowing agents with static capabilities to operate faster and more efficiently. I think that interactive flows may not benefit from this as much as just having a plethora of cli tools and a sandbox.
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.