An agent's most expensive mistake
When an AI operates a machine, the expensive mistake is rarely a misplaced click. It is believing something false about itself: that it has permission when it doesn't, that the machine is down when only the video is missing, or that a click didn't land and repeating it, until the form gets submitted twice. An agent that can't feel itself guesses, and guessing on someone else's computer is not acceptable.
Who am I and what may I do
The first tool any AI should use in NodalDesk is status. It tells the AI who
it is in the tunnel (its credential and trust level: full or limited), which
route it is on (direct or through the relay), what the machine on the other side looks
like (name, system, monitors) and, above all, what it is allowed to do:
browser, terminal, control, clipboard, audio, diagnostics.
If a permission is off, the AI knows before running into it, and asks the owner instead of
insisting. If video is unavailable, status says so: the machine is not down,
it is only missing its eye, and the AI can keep reading the screen through its structure.
Why something doesn't work
With the owner's permission, diagnostics shows the state of the machine's
capture, encoder, control, audio and route, with no secrets and no typing commands blindly.
It is the difference between "this is broken" and "I'm on limited trust, that's why I
can't move the mouse".
Did it really arrive?
Every keyboard or mouse action an AI takes gets an acknowledgement from the machine: whether it was taken, whether it was rejected and why, or whether there was no answer. When there is no answer, the result is "unknown", and NodalDesk never retries on its own: blindly repeating an action that may have landed is how a payment gets sent twice.
In the dedicated browser, browser_verify checks whether something really
happened (a text that appears, a button that goes away) without dumping the whole page,
and browser_watch reports as soon as something meaningful changes, instead of
looking at the page over and over.
Why it matters to whoever lets an AI in
An AI that knows where it is makes fewer mistakes, and when it can't do something, it says so. For the machine's owner that means fewer scares, permissions that are respected and an activity log that tells what happened, not what the agent believed happened.