Author here. I think the confusion is that “agent” is used for two very different things: - building an agent - an “agent” product/runtime (Claude Code, etc)

Author here. I think the confusion is that “agent” is used for two very different things: - building an agent - an “agent” product/runtime (Claude Code, etc) In the first case, the model never executes anything. It just outputs something like “call this API”. Your code is the one doing it, with whatever validation you want. There’s no need for a sandbox there because there’s no arbitrary execution.