CLI
The gateway through which a program is operated by text, without a screen. It’s what separates software that can be delegated from software that only accepts a human hand.
In one sentence
The command-line interface: you write what you want, the program does it and returns the result as text, with no window, no click, and no one watching.
Without it
Only has a screen — needs someone clicking
With it
Has a command — can be operated by another program
Analogy
Think of the loading dock at the back of the store. The front display was made for a person to look and choose. The dock was made for a truck to pull up, and it’s where the merchandise actually moves. It’s ugly because it serves a different purpose, and no one judges the store for that.
Example
The system that digitizes the letters only has a screen: to separate by language, someone opens it, looks, chooses from the menu, and clicks save. An agent has no hands, so that piece remains human, however capable the model may be.
The translation, on the other hand, happens on files in a folder, and there is a command that runs the entire queue. That piece the agent does end to end.
The speed difference between the two comes from one having a command and the other not. The AI model is the same in both cases.
The common mistake
Thinking CLI is “developer stuff” and ignoring it when buying software. In the Cognitive Enterprise, having a command line or API is a purchasing criterion, because it determines whether that process can someday be delegated or will stay stuck with a person clicking forever.
The opposite mistake is handing the terminal to the agent and walking away. A command runs with the permissions of whoever calls it. Without Sandbox and without Least Privilege / Permissioning, what you handed over was the house key.
In practice
- When evaluating a system, ask early: can this be operated without opening the window?
The answer changes that process’s score in prioritization.
- A good command fails loudly: nonzero exit code and a message that says
what happened. A command that fails silently poisons everything after it.
- Prefer commands that accept input and return output in plain text. That way
they fit into each other without glue.
- Log the exact command for each run, along with the result, so you can
repeat it and verify it.
How to make it tangible
A command that does the entire task end to end, and that anyone on the team can run by copying one line. If you have to open a window in the middle, it’s not ready yet.
Note
There’s an advantage almost no one notices: a click leaves no trail, while a command does. Every executed line becomes a line in the log, with what went in and what came out. That gives you half the Audit Trail / Provenance for free, which is why investigating what an agent did is easier than investigating what a person did on a screen.
Translated from Portuguese with AI assistance.