Not sure I follow you. If you're building a computer system where the "user" is intended to be another computer, why would you need anything besides an API? Or are you talking about a system where both humans and other computers are meant to consume it's services? If so, that would be somewhat interesting I guess.

But it seems to me that if you're going to use an LLM to "use" some other software, the way to go is use tool-calling support to call an API, and/or something like Anthropic's MCP (Model Context Protocol) stuff. There's some exiting work to, around "agent to agent" communications that one could use to integrate one kind of AI system with another computer system (whether or not the other system has any AI abilities). They range from things like FIPA-ACL, KQML, KIF, etc., through all the SemanticWeb standards, to some more recent specs that are being worked on. For example, the forthcoming ECMA TC56[1][2] standard for Natural Language communication between agents. And a similar'ish effort is mentioned in a recent arXiv paper[3].

[1]: https://ecma-international.org/technical-committees/tc56/

[2]: https://github.com/nlip-project

[3]: https://arxiv.org/abs/2411.05828v1