If the operating system is building the agent in, they will give it access to various APIs needed to do stuff. Apple already did this with Siri on the iPhone before AI took off, they wanted users to be able to adjust settings via Siri. Microsoft has made pretty much everything possible with Powershell as well, which I’d expect to be in Copilot’s training data. Maybe it’s not as easy as modifying a text file, but if the AI is doing it, does it matter?
Maybe it’s more work for the OS vendor, but it will likely be more user friendly overall, without any setup… this is the typical trade off and one reason why the year of the Linux desktop always seems 5 years away.
MacOS has quite excellent help docs built in. It also has all the "wiring" exposed via AppleScript for most applications.
I would suggest that the GNU+Linux command line is not accidentally designed well for agents, it is that agents are purposely designed to use command line. And there are far greater numbers of posts online about how to write a bash script to accomplish something with a command line than there are about similar tasks accomplished in AppleScript. The agents are designed for command lines, and they did not "learn" by using, they just "copy/paste with more steps" from the abundant chatter about the command line tools from the internet.
Maybe someone will write an agent in Java, instead of the sloppy JS that Claude Code is written in. Then you might get a cross-platform java.awt.Robot that can drive the UI. Maybe someone will write one in AppleScript and flex that scripting interface. Maybe someone will build one using a selenium driver for ChromeOS. Who knows.
The point is that one operating system has all the wiring exposed for all applications, and the others do not. And in some cases, they deliberately went the other way:
But Linux doesn't have wiring exposed for all applications on the command line, any more than any other OS. That isn't a requirement for applications to run on Linux. For example, the Eclipse IDE runs on Linux. It even has a command line interface. But what you can do in the command line is not everything you can do in Eclipse.
AppleScript arguably offers a far richer interface for driving applications that have a UI. It's just no one has designed an agent to take advantage of that.
If the operating system is building the agent in, they will give it access to various APIs needed to do stuff. Apple already did this with Siri on the iPhone before AI took off, they wanted users to be able to adjust settings via Siri. Microsoft has made pretty much everything possible with Powershell as well, which I’d expect to be in Copilot’s training data. Maybe it’s not as easy as modifying a text file, but if the AI is doing it, does it matter?
Maybe it’s more work for the OS vendor, but it will likely be more user friendly overall, without any setup… this is the typical trade off and one reason why the year of the Linux desktop always seems 5 years away.
Is it though? How did the agent learn to operate GNU+Linux? Man pages, or questions on forums, mailing lists, and wikis?
How does the answer to that question relate to the article?
MacOS has quite excellent help docs built in. It also has all the "wiring" exposed via AppleScript for most applications.
I would suggest that the GNU+Linux command line is not accidentally designed well for agents, it is that agents are purposely designed to use command line. And there are far greater numbers of posts online about how to write a bash script to accomplish something with a command line than there are about similar tasks accomplished in AppleScript. The agents are designed for command lines, and they did not "learn" by using, they just "copy/paste with more steps" from the abundant chatter about the command line tools from the internet.
Maybe someone will write an agent in Java, instead of the sloppy JS that Claude Code is written in. Then you might get a cross-platform java.awt.Robot that can drive the UI. Maybe someone will write one in AppleScript and flex that scripting interface. Maybe someone will build one using a selenium driver for ChromeOS. Who knows.
The point is that one operating system has all the wiring exposed for all applications, and the others do not. And in some cases, they deliberately went the other way:
https://www.jamf.com/blog/synthetic-reality/
But Linux doesn't have wiring exposed for all applications on the command line, any more than any other OS. That isn't a requirement for applications to run on Linux. For example, the Eclipse IDE runs on Linux. It even has a command line interface. But what you can do in the command line is not everything you can do in Eclipse.
AppleScript arguably offers a far richer interface for driving applications that have a UI. It's just no one has designed an agent to take advantage of that.