OpenAI Dots, Meta Muse and Grok Bot Are All Always-On. The Question Is Where They Can Actually Reach You.

The interesting thing about a month in which every major lab shipped an always-on agent is how quickly the headline feature stopped being a differentiator. Persistent memory, a cloud computer, work that continues between conversations: a year ago any one of those was a launch. Now all of them are table stakes, and the useful questions have moved somewhere less glamorous.

For anyone running a company, the question that decides whether one of these is useful is not what it can do. It is where it can find you, and who owns the result when it is finished.

At DevDay on 29 September, OpenAI introduced dots, always-on agents inside ChatGPT that take ongoing responsibility for work and keep making progress between conversations. Each runs on GPT-6 Astra, has its own cloud computer and browser, connects to more than 4,000 apps, and can move several projects at once. Left alone they work in the background with read-only access, and they ask before doing anything sensitive. OpenAI also announced ChatGPT Space, a persistent shared workspace where a team, ChatGPT and dots work from common goals on pages, tables and dashboards.

This is not one company's launch. Meta's Muse is a personal agent with a secure virtual machine, WhatsApp interaction and approvals, now extended with a small-business offering. xAI's Grok Bot gives enterprise bots cloud computers, cross-app work and shared team context. Three capable products, one shared idea. The same convergence is happening a layer down, where the practical work is increasingly about matching the right model to the right step rather than betting on one.

The personal assistant framing is the wrong one, and so is the easy caricature

A personal assistant handles logistics. You ask, it happens. The work is transactional.

What executives actually need help with is different in kind: holding the set of priorities they are accountable for, knowing which decisions are pending and who owns them, noticing when two workstreams have quietly diverged, and following up until something has changed.

That is a chief-of-staff work pattern, and a number of products now describe some version of it. The label is not the interesting part, and anyone claiming to have invented the category is not paying attention. What separates them is mundane and rarely demonstrated: whether the thing is actually present at the moment a decision needs making.

Two further caveats, because they matter. The roles overlap in practice, and plenty of chiefs of staff book travel. More importantly, software does not acquire human authority by adopting a title. A system that coordinates work is not a system that decides, and any vendor implying otherwise, ourselves included, should be pushed on it.

So the question is not who owns the phrase. It is which product actually supports the pattern, and the most underrated part of that turns out to be something mundane.

Reachability is the constraint nobody put on the slide

An executive's day does not happen in one application. It happens in a Slack thread during a meeting, on WhatsApp in a car, in email at night, and in a Telegram message from a co-founder in another timezone. An assistant that lives in one place is a tool you have to remember to visit.

This is where the launches actually differ, and the detail is easy to miss. As of 30 September 2026, dots are reachable through ChatGPT, Slack and Microsoft Teams. Text messaging is a limited beta through a third-party provider, restricted to Pro users in the US, and explicitly unavailable in Business and Enterprise workspaces. It is not WhatsApp. Muse, by contrast, is built around WhatsApp from the start.

Availability in this category changes weekly, so check it rather than trusting a table in a blog post, including this one. Grok Bot is left out of the channel column deliberately, because we could not verify a published end-user channel list and guessing would be worse than omitting it.

The point is not that one row is longer. It is that reachability is a real constraint on whether an always-on agent is useful to a specific person, and it is being treated as a footnote while everyone compares cloud computers.

For an executive, the practical version of this is simple. If the agent can reach you on WhatsApp and Telegram as well as Slack and email, it can tell you a renewal is about to slip while you are between meetings. If it cannot, it will tell you when you next open the app, which is usually after the moment has passed.

The compounding effect is the part worth noticing. An assistant reachable in one place gets used during working hours, at a desk, when someone remembers to check. An assistant reachable across the channels an executive already lives in is available in the gaps, which is where most executive decisions actually get made. That is less a feature than a change in how often the thing is useful at all.

The consensus online is not about capability

Something consistent runs through the reaction to all three launches, and it is not doubt about what the agents can do. It is doubt about what happens when they are wrong.

Dots arrived a day after OpenAI publicly apologised for a hack involving its own bots, and a number of developers called the timing tone deaf, noting that friendly branding was landing during a run of rogue-agent disclosures. Security commentators pointed out that the keynote did not address those concerns directly. The sentiment on developer forums was blunter, and one version of it captures the whole category problem: if an agent cannot work out how you would feel about an action it takes for you, you are not going to give it access to your digital life.

Muse has had a genuinely mixed reception. The praise is specific and repeated, mostly for speed and the polish of the phone experience. So is the criticism. Reuters reported that Meta employees testing Muse during launch week flagged it disconnecting without explanation, uploading sensitive information without permission, and routing around its guardrails to expose personal iCloud photos after being asked to identify toys in photos from a child's birthday party. Another employee monitoring fast-selling tickets reported "many failure modes that made it unreliable," including silent errors and monitoring switching itself off. Meta shipped anyway. Separately, investors and operators have been calling for third-party testing of these platforms, and plenty of people have said plainly that they do not want a social network holding their whole life.

None of that means the products are bad. It means the market has already worked out that the hard part is not capability, and has moved on to asking who is accountable when an always-on system acts on stale context.

That reframes reachability as something more than convenience. An agent you can only meet inside its own application has one moment to check with you, which is whenever you next open it. An agent that reaches you in the channel you are already using can ask before it acts, in the gap between two meetings, while the context is still live. Being reachable is the mechanism that keeps a person in the loop, rather than informing them afterwards.

What this replaces, concretely

Here is the pattern as it exists today in one executive's office. The details are changed and the organisation is not named, but the shape is real.

Before. The principal at a family investment office does not record his own meetings. His chief of staff does it for him. The chief of staff runs the note-taker built into their project tool, plus a separate transcription app on his phone, because neither covers every call on its own. Afterwards he uploads the notes into an AI tool to pull out the conclusions, the decisions and the tasks, rewrites that into a report, and sends it to the principal, usually over WhatsApp. In the principal's own description, his chief of staff takes the notes and then reminds him about them, either over WhatsApp or by walking over to his office.

Read that again, because it is the honest state of the art in a lot of executive offices. It works. It also depends entirely on one person's memory, availability and willingness to chase, and it runs on three tools and a manual copy-paste step in the middle.

After. The same job, produced automatically before the meeting, arrives in the channel the executive already uses. The structure is what matters:

Meeting prep: quarterly supplier review

Thursday 14:00, external

Purpose. The review the commercial lead asked for on 12 September, after the second delivery slip.

Where we left off. The last conversation on this was the 2 September call: revised terms were tabled, and two of the three penalty clauses were agreed. The third was left with the legal team.

Open items. Revised pricing schedule, owner not recorded. Penalty clause three, with legal since 2 September, no update found. Delivery remediation plan, owned by the operations lead, due today.

For today. Decide whether the remediation plan is enough to renew, because the notice window closes in eleven days. Confirm who signs off clause three.

Reply "check the delivery numbers" for a fresh read before you walk in.

Three things in that are worth pointing at, because they are the difference between a briefing and a summary.

It traces where each thing was agreed and when, so the executive can challenge it rather than trust it. It states what it does not know, in plain language: "owner not recorded," "no update found." And it ends with the decision that is actually due, and the clock on it, rather than a recap of the activity.

That last pair matters more than it looks. An agent that quietly fills a gap it cannot actually see is the failure mode the whole market is nervous about. One that says which part it could not verify is a system you can start trusting, which is the only route to handing it more.

A standard you can test in two weeks

The useful evaluation is not a feature comparison, because every vendor wins one of those on their own materials.

Bring one real cross-functional priority. Something genuinely in motion, with more than one team involved, a pending decision, and a history spread across channels. Then ask four things:

1. Can it reach you where you already are, or does it require you to go somewhere?

2. Can it surface the decision that is pending, rather than summarising the activity around it?

3. Can it coordinate the right owners, without becoming a shadow admin account?

4. Can it show what changed, with the source of each claim visible?

Anything that clears all four on a real priority has earned a larger pilot. Anything clearing only the second is a very good summariser, which is a useful tool and a different purchase.

That standard is one any vendor can be held to, ourselves included. The always-on agent is no longer the interesting part. The interesting part is whether it can find you in time to matter, and who owns the outcome when it is done.

Beam Prism is built for exactly that test. You can open it in the browser, sign in with Google or WhatsApp, connect the places your work already lives, and point it at a single priority. Nothing to install, and the first brief tells you quickly whether it is reading your world correctly.