Automations lives at /workbench/text/auto. It is the part of x-rush where a Character you own writes for you on a schedule: you describe the recurring task once, pick a time, and from then on each run produces a message that lands straight in that Character's AI Chat history, tagged as an automatic run so you can tell it apart from ordinary chat. Two things to unlearn first. Nothing here is on by default — an account with no Automations gets zero pushes, no matter how long it sits. And there is no roster of built-in agents to subscribe to. The writer is always one of your Characters, and one Character carries exactly one Automation.
The page has two tabs plus one public side door. History lists everything a run has delivered. Tasks is where Automations get built, edited, paused, and deleted. The side door is Explore, the community shelf where people publish task configurations for others to remix. If you only read one tab, make it History; nothing appears there until you have done the work in Tasks.
Start With a Character
Every Automation is bound to one Character, and the binding is the entry requirement. Open Tasks in a Project with no eligible Character and the empty state tells you plainly: create or copy one in AI Chat first. Automations has no inline Character editor. You build the persona, its background material, and its chat history where that persona actually lives, then come back. Built-in Characters in the Default Project can bind without copying. In any other Project the Character has to already live there.
The pairing is exclusive and permanent, in both directions. A Character already bound to a task disappears from the picker. After you confirm a binding you cannot swap Characters later; if a different persona should run the schedule, you create a new Automation for it. What does follow you for free: edit the Character in AI Chat and the next run inherits the change, no re-confirmation, no extra cost.
Configure, Preview, Enable
Creation runs through a fixed six-step form: Basic information (name, Character, optional tags), Task requirements (the recurring goal, plus must-include and must-avoid notes), Web search (an explicit allow or deny, with preferred sources, excluded sources, and a citations toggle), Output controls (language, length, format, fixed sections, all defaulting to the Character's own habits), Schedule (one execution time, one explicit time zone), and Final review. A Workflow Copilot can draft suggestions into the same form, but it cannot enable anything, and each suggestion it successfully drafts bills like one ordinary chat answer — the Automation's own fixed price is untouched.
My own setup: a briefing Character whose goal I filled in as "Every weekday at 07:30, write a three-item morning briefing: one training focus, one science story, one dinner suggestion." Seven-thirty because my gym opens at eight. The schedule field is deliberately narrow — every day, weekdays, or a custom pick of weekdays, one time slot, one time zone that stays pinned after you save. The same task twice a day means two Automations and two Characters.
Fill in the name and the first requirement and the draft saves itself: it shows up in Tasks, occupies the Character's binding, and schedules nothing. Before anything runs you confirm a free configuration preview — the complete form, the next execution time, and the fixed credit price per real run as quoted by the server. Opening the preview generates nothing and spends nothing. From there you can enable directly (enabling is free) or fire a paid trial first. A trial is a real run: same fixed price, output delivered to History marked as a trial, and it enables nothing by itself. Enabled or not, nothing runs at save time; the first delivery arrives at your scheduled slot.
History: Where Delivered Content Lands
History lists successful deliveries only. A run that failed, hit a safety block, or got skipped stays out of this list — those live in Run History under the task. Each entry indexes the same message that sits in the Character's AI Chat history, so opening it from either side shows the same text. Badges separate scheduled deliveries from trials and manual reruns. Unread entries keep their marker until you actually open one, and a delivery carries its web sources only if the task's citations toggle was on when it was created.
Two actions matter here. Continue in AI Chat jumps to the Character and scrolls to that exact message, so your follow-up lands in the same thread; nothing is re-inserted or regenerated. Run again starts a fresh paid run with the current configuration — priced before it starts, and the original message stays untouched.
Run History: The Execution Ledger
Open a task's details and you get the ledger: one immutable row per run, filterable by status and time. In progress, succeeded, system failed, balance required, content safety block, cancelled, skipped. Three behaviors are worth knowing before they surprise you.
If your credits run out, the run does not execute and the row says so; after the first shortfall the Automation pauses itself, and resuming asks you to choose between waiting for the next slot and running once immediately. After three consecutive failed runs it pauses too — any success resets the count. And if a run is still going when the next scheduled time arrives, the new cycle is skipped and logged; nothing queues and nothing double-bills. A system failure gets one free internal retry before the row records a failure. Your own retry on a failed run is a new paid run, quoted up front.
Deleting: What Stops and What Stays
The delete confirmation spells out the deal: stop all future execution, permanently delete the configuration, release the Character binding. Then the line most people don't expect: "Existing successful messages remain." I had assumed deletion would sweep the history with it. It does not, and that is deliberate — the delivered messages belong to the Character's chat history, and the task config was only ever the schedule that produced them. Deleting kills the configuration for good (no trash bin, no restore) and frees the Character, so the same Character can take a new Automation later; its history, automated messages included, stays exactly where it was. Credits already spent stay spent.
Related lifetimes, briefly. Clearing the Character's chat history in AI Chat removes the automated messages too, while the Automation survives and keeps its schedule. Deleting a bound custom Character asks twice, then takes the Automation down with it. Neither refunds anything.
What Stays Private
Generated content stays private, full stop. There is no share link for a delivery, no public page, no community posting of outputs. The only thing that can go public is the task configuration itself: name, goal, schedule shape, search settings, output format, plus whichever parts of the Character's public setup you explicitly tick during publishing. That snapshot lands in Explore, where anyone can remix it into their own Project as a fresh Character and a fresh Automation starting from an empty chat history. Your messages never travel with it. Delete the source Automation and its Explore entry goes down with it; remixes other people already made keep running as their own copies.
The Quick Chat Bubble
Once an Automation is enabled, its Character appears in the site-wide quick chat bubble, everywhere you are signed in. The badge counts unread automated messages across all your Projects; pause or delete the task and the Character drops out of the bubble immediately. Inside, you get the message list, the full text, and a reply box that writes into the same Character history — those replies bill as ordinary chat. Delivered messages never pop open on their own. You click; they open.
FAQ
I enabled a task and nothing arrived. Open the ledger. A row marked balance required means top up and resume. A row marked skipped means the previous run was still going when this slot came due. No row at all means the time has not come yet — and the schedule's saved time zone decides, not your laptop's.
Can one Character run two schedules? No. One Character, one Automation, and the binding cannot be swapped. Copy the Character in AI Chat and give the copy its own Automation.
Can I share what it writes? Not through any public mechanism. A delivery is visible in your History, the Character's chat history, and the bubble, and nowhere else. Explore publishes configurations, never outputs.
That is the whole surface: bind a Character, fill six steps, confirm a free preview, and the schedule takes over. The boundaries are worth remembering. Runs bill the fixed quoted price and no configuration action does; deletion stops the future and keeps the past; and everything the Character writes stays yours.