OK, but isn't this just programmatic tool calling, which is beneficial in some cases but not others?

This is available in DSH via PTC mode.

Basically, instead of the completion API returning a sequence of tool calls, it returns a program that invokes the tools.

I think most are returning Typescript program

With ordinary tool calling, you may have to wait for a tool to return its result to the completion API before the model can continue and issue the next tool call.

With PTC, the generated program can invoke independent tools asynchronously, collect their results, and dispatch the relevant results back to the completion API in a single turn.

Claude and Codex added this as well, but there are some papers on this that show it's not good across all tasks, and models need to be trained for this specific technique, which is what AI labs are doing now.