This lifts the work to be done from “what is the proper api to show a modal with this framework” to “what does it mean t...

This lifts the work to be done from “what is the proper api to show a modal with this framework” to “what does it mean to block user input? What is a hardware jobs that is fine with running while user input is held up? Does it matter? Do users ever need to interact with ongoing work when a warning is shown?”, things where our more formalized language is much less precise/consistent than “modal”.But maybe we can transform “modal.confirm(X)” to “userInputBlockingConfirmationRequest”, and that makes it possible to prompt the llm with “when a file can’t be save, block user input and request confirmation”, but we lose the “warning modal” words. Etc…This is how I approach llms and it has nothing to do with agents. Having harnesses are nice, because they make it possible for a model to discover that a modal is a blocking confirmation request, or write files, etc…But that’s compiler work vs language work, and comparatively, compiler code is fairly straightforward once language work has been wo...