Automations
Run an agent prompt on a schedule and send the result to a linked chat
An automation is a prompt that an agent runs on a schedule. Use it to check, summarize or report on something without asking each time. Open Agents, select an agent, then select Manage and Automations.
Create an automation
Select New automation and fill in the form.
| Field | Options |
|---|---|
| Name, Prompt | The prompt is the task the agent runs each time. |
| Schedule | Daily, Weekdays, Weekly, Hourly or Cron (five fields: minute hour day month weekday), with a time zone. |
| Conversation | Same conversation continues one conversation, so the agent remembers earlier runs. New each run starts clean and keeps the latest 20 conversations. |
| Model | Agent default, or another model for this automation. |
| Deliver to | Web app only, or a linked chat from any channel. |
| What to send | Final answer sends the finished answer to the chat. Only what the agent notifies sends nothing unless the agent decides to notify you. |
Use Only what the agent notifies for checks that need your attention only sometimes. Write the prompt so that the agent knows when to notify, for example "Notify me only if a pull request waits for my review."
The agent can also create and change automations itself. Ask it in a session or in a chat. An automation that the agent creates from a chat sends its results to that chat. One created from the web app sends results only to the web app.
Manage automations
Open an automation to Run now, Edit, Pause, Resume or Delete it. Deleting an automation also deletes its run history.
Recent runs lists each run with its status, the prompt it used, its answer and its delivery result. Open in conversation opens the session of the run. If you delete that session, the answer of the run is no longer available.
An automation is never paused automatically. If runs fail, the automation shows how many failed in a row and keeps its schedule.
Skipped runs
A scheduled run is skipped, and recorded as Skipped, when:
- The previous run is still going in the same conversation.
- The agent service is not ready.
- The agent service was asleep or down for more than 15 minutes past the scheduled time.
A skipped run does not run later. The next scheduled time runs as usual.
Delivery to a chat
The agent checks the chat before it sends. Nothing is sent when the chat was unlinked or linked again after you created the automation. The run then shows Could not send. Edit the automation and choose the chat again.
A final answer is sent at most once. If the service stops while it sends, the answer can be lost, but it is never sent twice. The answer always stays in the web app.
Each run is a normal agent task. It spends your balance through the agent key at the prices of the model it uses. TokenFlux does not limit how many automations an agent has or how often they run.