Agent notifications
Let an agent tell you when something needs attention — and understand why sometimes the notice does not appear.
An agent that finishes a long job, or that needs a decision, has a way to tell you: it asks for a desktop notification.
The command
elyra notify "Tests are green, ready for review"elyra notify "Needs a decision on the migration" --title "orq-2-4"The text is the body of the notice; --title is optional. The command fires the notification and moves on —
it waits for nothing.
Why the notification sometimes does not appear
An agent that warns and is not heard is worse than an agent that does not warn, because whoever is waiting thinks the silence means “still working”. The most common reasons for suppression:
| Reason | What happened |
|---|---|
| You are already looking at that terminal | The notice does not appear on purpose: you are there, the notice would be noise |
| Text repeated within a few seconds | The same text is delivered only once; nearby repetitions are discarded |
| Notifications turned off | The general alerts switch is off |
| That source’s channel is off | Alerts have per-source control; the agent’s is the general switch |
| Nothing changed, but the window is hidden | A minimized window lights the attention dot in the tray instead of showing the banner |
| No desktop to notify | In an environment without a graphical session, there is nowhere to show the notification |
The attention dot in the tray exists because a banner you do not see warns nothing — and it clears when the window comes back.
A design decision worth knowing
An agent notification has no switch of its own. An agent only notifies when someone told it to — it does not decide to warn on its own. That is why the control is the general switch: off, no agent notice gets through.
At the same time, the agent shares the same repetition table with the notices the app itself generates (terminals and task completion). This avoids two notices for the same moment — one from the app saying the agent finished and another from the agent saying the same thing.
How to use it well
- Warn about what requires action. “I finished” is useful when someone is waiting; “I am working” is not.
- Write the text thinking about what you will read two hours from now, when you will not remember the context.
- Do not repeat the same notice. It will be discarded, and you will not know whether it was suppression or whether the agent got stuck.
- One notice per subject. Two important subjects in the same text become one notice that you will reopen and reread.
Scope and limits
- The notice carries no context. It says what the agent wrote; opening the terminal is what gives the context.
- A notification is not a record. It is ephemeral: what needs to last goes to the conversation, to the board or to the task comment.