ChatGPT Dots vs Meta Muse: The Always-On Agent Decision

The next AI decision is not simply which assistant gives better answers. It is which assistant should hold a responsibility after you stop watching. That is the proposition behind Meta Muse and ChatGPT Dots: a persistent agent that knows enough about your goals to move work forward between conversations.

Meta announced Muse on September 8. OpenAI introduced Dots on September 29. Both give the agent its own cloud computer, both can maintain context, and both describe proactive help. Yet choosing between them as if they were two equivalent chatbots misses the practical question: which one fits the work you can safely delegate?^{[1] [2]}

Our view is that these products compete on the cost of supervision as much as intelligence. Muse starts closer to a personal concierge. Dots starts closer to a continuing work relationship inside ChatGPT and its connected tools. Neither description is a hard capability boundary, and neither company has established that its agent can reliably own every responsibility it advertises.

This comparison draws on public product documentation and independent reporting. It does not present a controlled Intueo performance benchmark of either agent.

Click to expand

Similar promise, different starting point

Muse runs on Muse Secure VM, a dedicated virtual computer with a browser and stored information for the individual user. You can talk to it in its apps or through WhatsApp. Meta's launch examples include travel, shopping, bill negotiation, and plans that adapt as someone's circumstances change.^[1]

Dots is powered by GPT-6 Astra and starts with relevant ChatGPT memory.^[6] It uses connected tools and Codex to research, prepare documents, analyze data, and build software. You can reach the same dot through ChatGPT, Slack, Teams, and calls, while its saved notes and relevant context carry forward.^{[2] [3]}

Keep the names straight. Muse is the personal-agent product. Muse Spark is Meta's underlying model family. Muse Code is a separate coding product. Dots is an agent product powered by Astra, not another name for the model itself. A leaderboard result for Spark or Astra cannot, on its own, tell us which complete agent will be easier to supervise.

An independent September 30 comparison in Tom's Guide also framed Muse as more accessible for everyday users and Dots as more closely aimed at paid professional workflows. That is outside editorial judgment, not an Intueo performance finding. Where a review's rollout details differ from current official documentation, the official access requirements should govern.^[7]

The cloud matters more than the word always

Both agents can continue work after the app closes. Dots documentation explicitly says its cloud work can continue when your computer is off. Muse's dedicated cloud VM similarly means its main environment is not your laptop.^{[1] [3]}

But cloud work and work on your connected personal computer are different. For Dots to use your local files, apps, or local skills, that computer must be online with the ChatGPT app open. Its documentation allows one connected personal computer at a time. Connecting Slack does not also connect your inbox, other apps, or laptop. The cloud browser also has its own sign-in sessions rather than inheriting your local browser's accounts.^[3]

This distinction can determine whether a promised workflow is realistic. An agent may research public suppliers overnight on its cloud computer, but it cannot necessarily inspect a spreadsheet locked on a sleeping workstation. A private application may block cloud browsers or require a verification step only you can complete.

Always-on therefore means persistence and background availability, not a guarantee of uninterrupted execution or successful completion. Dots can pause and wake up as work requires. Muse returns when something changes or approval is needed. Access failures, usage allowances, missing information, and review checkpoints still shape progress.

A useful brief specifies not just the goal but the signals to watch, where progress should appear, and when to stop. Otherwise an agent may spend attention on work that is technically relevant but no longer important. Persistence is valuable only when it follows current priorities.

Memory should reduce repetition, not scrutiny

A good personal agent should remember a preference without making you repeat it every Tuesday. Muse says users can tell it to forget specific things, and its technical description says users can inspect, edit, and download its memory files. Dots uses relevant ChatGPT memory and its own saved notes about preferences, decisions, and ongoing work.^{[1] [3] [4]}

That helps with continuity, but memory is not a guarantee of truth. An old preference can become a bad instruction after a budget change, a new client, or a change in family circumstances. A remembered fact should sometimes trigger a question rather than an action.

Consider a fictional freelance photographer, Lena, planning an exhibition. She wants the agent to compare print suppliers, maintain a delivery checklist, and draft answers to venue questions. She does not want it to accept a contract, increase the print budget, or email guests without review. This is a fresh illustration, not a customer case study.

Muse might fit if her task is mainly shopping, bookings, and personal coordination. Dots might fit if her files, feedback, and ongoing decisions already live in ChatGPT and connected work tools. In either case, the helpful memory is not simply that Lena prefers matte prints. It is also that the budget is provisional and the venue has not approved the dimensions.

The permissions checkpoint

Permission is not one yes-or-no switch. There are at least three distinct decisions: what the agent can read, what it can prepare, and what it can commit in the outside world. Reading a bill is different from paying it. Drafting a contract summary is different from signing. Preparing an email is different from sending it.

Muse documents a separate Sentinel that is the permission authority for connector actions and network traffic. The main agent proposes actions; Sentinel evaluates whether they are allowed, denied, or require a user decision. Credentials are stored separately, and real secrets can be inserted at the network boundary without showing them to the main agent.^[4]

Meta also describes separating read and write permissions where supported, finer controls for connector actions, and approvals scoped to particular uses. Its purchase flow asks for approval and can use a single-use payment credential. This is a documented architecture intended to limit damage, not proof that mistakes or prompt injection have been eliminated.

Dots combines existing app permissions, built-in safeguards, Custom Rules, and Auto-review. Before actions affect accounts or share information, automatic review determines whether work may proceed, needs approval, or must be handed over. Password changes are an example of a step that stays with the user.^{[2] [5]}

The distinction between unprompted and assigned work is important. Dots' proactive research uses read-only tools. That does not make every task it runs read-only: authorized follow-up work can take actions within its applicable permissions. Custom Rules can require approval or handoff, but OpenAI says they are instructions the dot tries to follow and can make mistakes. They do not grant app access or override built-in safety requirements.^[5]

For Lena, start with narrow reading access, draft-only communication, and human checkpoints for money, commitments, or publication. Expand authority after reviewing repeated results. Least privilege makes experimentation affordable.

Privacy promises need their own comparison

Muse Secure VM is the launch architecture. Muse Confidential VM is a different, stronger privacy mode planned for later in 2026, with a small trusted testing group described in Meta's technical post. Do not buy the current product assuming the later mode is generally available.^{[1] [4]}

Meta explicitly says its current architecture restricts personnel access through operational policies but does not prevent Meta from accessing data when necessary to operate, secure, or support the service. Its promised confidential mode aims to cryptographically prevent that access. Those are materially different assurances.

Muse says it does not share conversations or VM data with Meta's ad systems. That is not the same as saying nothing the agent does can influence advertising. Visits to merchant websites can still have downstream advertising effects. Meta also describes sanitized inference trajectories being used for model training, with an opt-out in Muse settings.^[4]

OpenAI says Business, Enterprise, and Edu workspace content is not used to improve models by default. Personal-plan users can control use of eligible Dots conversations and work through data settings. It says it does not train directly on proactive research or private notes, while information from them can enter eligible conversations or tasks and then follow the relevant settings.^[2]

Neither policy should be simplified into an absolute privacy slogan. Before connecting sensitive business records, verify the applicable plan, organizational requirements, and the precise permissions being requested. Do not paste credentials into a conversation merely because the agent needs an account. Use the documented private sign-in or connection flow.

Access and the cost of a useful trial

Muse's September launch describes a US rollout on iOS, Android, and muse.ai, free for most needs with subscriptions for heavier use. That is not evidence of universal international availability or unlimited execution.^[1]

Dots is rolling out gradually. The official documentation lists Pro 100, Pro 200, and Pro 500 for adults outside the European Economic Area, United Kingdom, and Switzerland. Business Premium and Enterprise are rolling out worldwide; Enterprise access is off by default and requires an administrator. This is not an all-Plus benefit or instant access in every account.^[3]

OpenAI says the first dot is included in eligible Pro or Business Premium plans, with an allowance for deeper work. Conversations do not count toward ChatGPT usage limits, but tasks started or managed in Work or Codex count toward those products' usual limits. Included does not mean unlimited.^[2]

Finally, learn how to stop work before expanding it. Dots' Pause stops the main task, not every delegated task or future schedule. Those need separate controls. Stopping an agent also does not undo a sent message or completed change.^[5]

Our verdict is to choose the smallest responsibility that produces a measurable benefit. Muse is a plausible first trial for US personal administration. Dots is a plausible fit for eligible users whose work already runs through ChatGPT and connected team tools. The better agent is the one that returns usable progress with fewer corrections and fewer unnecessary interruptions, while leaving consequential judgment where it belongs: with you.