You've got a task that already works in chat. The next tempting step is to let the agent keep going after you close the laptop. Grok Bot can do that: the FAQ says Bot work runs on the cloud computer, and closing the app, laptop or phone does not stop a background turn or a routine.
What changes overnight is not the computer. It's that you won't be in the chat. An approval card raised by a routine or other background work expires after 30 minutes, and the action doesn't run. So before you set the schedule, decide what the agent may do alone, what must wait for you, and what is off-limits.
In earlier articles I covered the first brief and how to turn a reliable run into a routine. This one is the permission decision those two assume you've already made. Grok Bot behaviour below is from the official docs, fetched on 18 Nov 2026. The three lists are my checklist, not a built-in setting.
Write three lists before you set the schedule
The approvals and security page tells you to start with read-only tasks and draft outputs. It also lists the actions that should have an explicit boundary in the request:
- sending messages or invitations
- publishing content
- purchases and financial transfers
- deleting or overwriting data
- changing permissions
- production changes
- accepting legal terms
I split the job into three lists, plus connectors, a morning report, a missing-source rule and a stop rule. For a research digest that should only draft:
job: Overnight research digest
schedule: "Every weekday at 07:00, Europe/Rome"
may_run_alone:
- Read today's official documentation pages
- Draft a digest in this conversation, with source links
- Flag items that need a decision
ask_first:
- Send messages or invitations
- Publish content
- Post to Slack or any external channel
- Delete or overwrite data
off_limits:
- Purchases and financial transfers
- Change production systems
- Change permissions
- Accept legal terms
connectors:
- none (public documentation pages only)
report: Post the digest in this conversation, with source links and a short action log.
missing_source: Report the failure. Do not use yesterday's digest or remembered facts.
stop: If an action needs approval, wait. Do not retry a denied or expired card with a broader request.
Grok Bot does not read this file; paste the lists into the routine instruction. I run a small Python checker (standard library only) so I don't leave a consequential action in the "alone" column. Real output of python3 check_policy.py policy.yaml on that sample:
Warnings: 0
All seven official consequential actions are in ask_first or off_limits.
Required fields are present. No consequential action is in may_run_alone.
Zero warnings means the sample covers the official list, not that a live Bot has been tested against this file.
Put the boundary in the request, not only in settings
The docs start with a boundary inside the request: their example drafts a budget change, forbids changing the campaign or messaging the agency, and asks for approval after showing the current and proposed values.
Auto Review is on by default and checks actions before they run. Your Custom Rules can say Ask first or Allow automatically; if both match, Ask first wins. Write narrow rules ("Ask first before sending any external email"), not "allow everything in the browser." The docs are blunt: Auto Review is model-based and "should complement, not replace, least privilege and explicit approval boundaries."
The FAQ also says to tell each Bot its standing boundaries and ask it to remember them. That is useful, and it is not enough. The Bots page says memory is not a source of truth, and tells you to back safety boundaries with approvals and Auto Review rules rather than memory alone.
Connect only what the overnight job needs
The same security page has a least-privilege list: connect only the tools a workflow needs, use scoped service accounts where the source system supports them, and review connectors and routines regularly. Pause a routine when its source system changes.
All of your Bots share one cloud computer, so files, browser sessions and credentials on it are available across the roster. The docs say plainly not to use separate Bots as a security boundary. An extra connector is available to every Bot, not only the overnight one.
Two settings matter more once you're away:
- Local execution. Under Settings → Computer → Your computer, Execution on this computer defaults to Ask every time. The docs recommend Never allow unless a Bot has a reason to work on your local files.
- Payments. A card saved in Vault fills in at checkout without a separate approval each time. If you want to approve every payment, don't save the card.
Approvals work differently when you're not in the chat
I covered the full routine setup in the earlier article. The overnight details to keep in view are:
- In a chat you're in, an approval card waits for you. A card from a routine, a trigger or a message from another Bot expires after 30 minutes, then shows Expired, and the action doesn't run.
- Email from a routine, sent from the Bot's own address, waits as a draft until you press Send.
- Auto Review takes into account what you asked for in your own messages. Text from an email or a web page doesn't count as your request.
- The troubleshooting page says to turn on Notifications for Bots that run routines, so you get a system or mobile notice when the Bot is waiting for you.
- After a long period away, Grok Bot may ask whether to keep routines running and pause them if there is no response. Review paused routines when you return.
The 30-minute window is why "publish it while I sleep" doesn't belong on the "alone" list. If you miss the card, the publish doesn't happen — unless you've already allowed that class of action to run on its own.
Add a stop rule and a morning report
When you create a routine, the skills and routines page asks you to confirm six things: the owning Bot, the schedule and time zone, the input source, the expected result, the approval boundary, and what happens when a source is missing. The report and stop rules belong there too.
The report. The docs tell you to "Tell the Bot where to report partial completion," and to "Preserve source links and an action log for important decisions." The overview describes a morning briefing as one Bot summing up mail, calendar, Slack and docs in a single post, and the Gmail example leaves every draft for you to send.
The missing-source and stop rules. The docs' own routine example ends with "If the source data is unavailable, report the failure instead of using old data." If an action needs approval, wait. Don't retry a denied or expired card with a broader request.
A new routine waits for its next scheduled time. Use Test only with safe inputs: a test run performs real work.
Checklist
- Write the three lists (alone / ask first / off-limits) before you set a schedule.
- Cover all seven official consequential actions in ask first or off-limits.
- Put the same boundary in the routine instruction, not only in Auto Review rules.
- Connect only the apps the job needs; review connectors regularly.
- Don't treat a second Bot as a security boundary: they share one computer.
- Leave local-computer execution at Ask every time or Never allow unless you have a reason; don't save a payment card in Vault if you want to approve every payment.
- Turn on Notifications for Bots that run routines.
- State the morning report, the missing-source rule and the stop rule, then Test with safe inputs.
Where to go next
The free Grok Bot QuickStart: Your First AI Teammate in 30 Minutes is a 7-page PDF with a 4-question test, five setup steps and a brief template, the trust ladder, the six things every routine must state, three starter briefs and five beginner mistakes to avoid.
The full path is in Grok Bot in Practice: Build an AI Team That Works While You Sleep, a 53-page guide covering connectors, skills, routines, approvals and Auto Review, a coordinated team, the weekly usage allowance and common failures, with seven case studies, 10 Bot briefs, 5 routine recipes and a 30-day plan.
This article was written with the assistance of AI and reviewed by the author. Grok Bot behaviour is described from xAI's official documentation (approvals, security and privacy; skills and routines; FAQ; Bots; settings and notifications; troubleshooting; and overview), all fetched on 18 Nov 2026. The checker output is the script's real output on the sample policy above, not on a live Bot run. PromptMaster guides are independent and not affiliated with or endorsed by xAI, SpaceX or Cursor.
Top comments (0)