Grok Bot and OpenAI’s dots: always-on agents, and what to check before you hand one the keys
Published
Two of the big AI labs launched the same idea within seven weeks of each other. On 11 August 2026 xAI released Grok Bot in beta, and on 29 September OpenAI introduced dots at DevDay. Both are agents that keep working when you close the chat window: you give them a goal, they work across your apps on their own, and they come back when they need a decision.
That is a different thing from a chatbot, and it deserves a different kind of care.
What Grok Bot is
xAI describes Grok Bot as always-on teammates. Each Bot works on a cloud computer, signs into the tools you already use and moves across apps, inboxes and websites to finish multi-step work. You message Bots like colleagues, they remember earlier conversations, you can teach them a workflow once and save it as a routine, and they can coordinate with each other.
At launch it was in beta on desktop and iPhone for SuperGrok and Cursor subscribers, with enterprise access by waitlist. Bots have their own usage allowance, separate from the Grok or Cursor plan.
What OpenAI’s dots are
OpenAI’s dots are personal agents built on GPT-6 Astra. A dot pursues goals you set in the background, and you can reach it through Slack, Teams and other workplace tools, with text messages to follow. You start with one primary dot; teams of dots and “specialist” dots with their own identities and credentials are the next step. At launch they were for ChatGPT Pro and Business Premium users in eligible markets, launched from ChatGPT or Codex, and OpenAI is working with Microsoft to bring them under Agent 365’s security controls.
How they ask permission
This is the part worth reading closely, because it decides what can go wrong.
xAI’s documentation is unusually specific. An automatic review checks each action first; most run without asking, and the rest show an approval card where you allow once, always allow or deny. Approval cards raised by background routines expire after 30 minutes. Emails a Bot writes from a routine wait as drafts until a person presses Send. And two warnings stand out: all of an account’s Bots share one cloud computer, including browser sessions and saved command-line credentials, so separate Bots are not a security boundary; and saved payment cards fill in at checkout without a separate approval unless you leave that option off.
For dots, the public detail on approvals is thinner so far. The launch coverage centers on what they can do and on the Microsoft security integration that is still being built.
What we would set up before handing one the keys
None of this is a reason to stay away. It is a reason to treat an always-on agent like a new hire with a company login, not like a search box.
- Give it its own accounts. A dedicated mailbox, a dedicated user in your CRM and a separate payment card with a low limit. If the agent shares your personal sessions, every mistake happens with your name on it.
- Decide what always needs a person. Sending to customers, paying, deleting and changing anything in finance should wait for an approval. Reading, drafting and summarizing usually don’t need to.
- Turn off what you don’t need. Saved payment cards, local computer access and broad “always allow” rules can all be switched off until there is a reason to switch them on.
- Keep a record. Know where to see what the agent did yesterday, and who approved it. If the product doesn’t make that easy, log it yourself.
- Start with one routine. Pick a task that is frequent, low-risk and easy to check, like collecting weekly figures into a report, and watch it for two weeks before giving it more.
These are the same rules we build into the automations we deliver: rules first, a person at the moments that matter, and a log of every run. An agent that can sign into anything makes them more important, not less.
Where this leaves workflow automation
Always-on agents are good at work that changes shape every time: research, triage, the odd one-off task across five tools. Fixed, high-volume processes such as invoicing, order sync and lead routing are still cheaper, faster and easier to audit as plain workflows that run the same way every time. In practice the two meet in the middle: a workflow does the predictable part, and an agent handles the exceptions it hands over.
Sources: Introducing Grok Bot (xAI, 11 Aug 2026), Grok Bot approvals, security and privacy (xAI documentation), OpenAI launches dots, its bubbly agentic avatar (TechCrunch, 29 Sep 2026).
