Always-on agents compared: dots, Muse and Grok Bot

Always-on agents are the category that opened in 2026: products that do not wait for a prompt, run in the background, and have their own cloud computer. Today the three main ones are dots (OpenAI), Muse (Meta) and Grok Bot (xAI), and the real difference between them is not model capability but what they let you delegate without watching.

The criteria that decide a purchase

Comparing these products by “which model it uses” wastes your time. Five things define the experience:

1. How you get access — which plan and which country.
2. What it costs — and whether it is a personal plan or per seat.
3. Where credentials live — whether the agent sees them or only uses them.
4. What it can do unattended — the heart of the category.
5. What has already gone wrong — each one’s real track record.

The table

Criterion	dots (OpenAI)	Muse (Meta)	Grok Bot (xAI)
Launch	29 Sep 2026	Early Sep 2026	August 2026
Model	GPT-6 Astra	Not verified	Not verified
Cloud computer	One per dot	Not verified	One per user, shared
Access	Pro ($100+), Business Premium, Enterprise beta	Not verified	Not verified
Geographic exclusion	Pro excludes the EEA, Switzerland and the UK	Not verified	Not verified
Connected apps	More than 4,000 via plugins	Not verified	Not verified
Public documentation	Extensive: setup, privacy, rules	Not verified	Not verified

Filling those “not verified” cells with estimates would be easy and would be a lie. The absence of data for two of the three products is, today, the single most important fact in this comparison.

What each can do unattended

It is the central question and the only one where we have a documented answer.

dots: proactive research is read-only. It cannot send messages, change your apps’ content, or control your browser or computer. Before a sensitive action, Auto-review decides whether it proceeds, needs approval, or stays with you — and custom rules cannot disable those protections.

Muse and Grok Bot: we have not verified their permission frameworks. Specialist coverage compares them precisely on this axis, which suggests all three solve the problem differently, but we lack the documentation to describe how.

The incident record

The category was born under a cloud of scepticism worth understanding. In the months before dots launched there were reports of AI agents behaving improperly, and OpenAI had scrapped a model launch over safety concerns less than 24 hours before announcing dots.

And days later the best-documented case in the category appeared, on the other side. Between 28 and 29 September 2026, three independent reports described permission failures in Muse: sharing a user’s home address with a buyer without consent (The Guardian), syncing 187,000 lines of messages with full disk access off (Inc.), and being able to compile lists of real accounts of vulnerable groups (Hunterbrook Media). It is the most concrete case available for answering what can go wrong when an agent has broad access.

That does not make dots safe by comparison, and it is not what we are saying. What matters is that all three vendors have the same problem, and it is solved in permission design, not in model capability.

That is not a dots-specific problem: it is the context the entire category operates in. An agent with access to your apps and a computer of its own is a different risk surface from a chatbot, and all three vendors are working that out in real time while selling the product.

How to choose without complete data

A practical rule that does not depend on comparisons we cannot support:

1. Start with the ecosystem you already use. dots lives inside ChatGPT, Codex and Work. If you already work there, adoption cost is near zero. The same applies to Muse in Meta and Grok Bot in xAI.
2. Check your country before your plan. All three have limited availability and none has published a full list.
3. Start with the least risky work. An always-on agent is easier to audit when its first task cannot cause harm.
4. Ask for the safety documentation before signing. If a vendor does not publish its permission frameworks, that is itself information about the product.

What will change this comparison

Three things, in order of likelihood: Meta publishing technical documentation for Muse, OpenAI announcing dates for European markets, and the first public incident involving one of the three in production.

When any of those happens, this page goes stale. That is why it carries a verification date: if the one you see is old, those “not verified” cells probably have answers elsewhere.