This is a paradigm I’ve been using well before all this tool nonsense.

You have a file which defines a series of typed functions. You ask the LLM to generate code for a task given those functions and then send* that code to an executor which runs the code.

If you need tool calls you make those inside of functions which call an own LLM with that capability.

You can do a lot of crazy things this way, plus you can very easily audit what’s happening. You also can have very specific contexts tied to nested LLM calls.

I’m open to arguments that tariffs can be effective, but their advocates rarely seem open to the argument that the way they’re implemented matters.

To put it kindly a lot them are chaos apologists.

I’d guess Run Forest Run.

At the same time, that review concedes he “would bet the number is more like three felonies a month.” That seems sufficient for abuse.