OpenAI dots vs Muse vs Grok Bot vs Instinct: personal agents

On this page

Personal AI agents are moving into the mainstream. OpenClaw and Hermes helped show what an assistant could become when it had memory, tools and time to work between conversations. Meta, xAI and OpenAI are now putting their versions into familiar apps and subscription products. Instinct takes another approach: an assistant you can text or call.

We've read the product launches, setup guides, security documentation and terms to understand what each company is offering. We compared where the agents work, whose context they carry, when they need permission and what access includes. We also traced the model back to OpenClaw and Hermes. The products differ in how they organise work: one agent working for a person, several Bots working together, or an agent shared by a team. Their terms still put responsibility for delegated actions on the user. We haven't tested their performance against one another.

Key takeaways

• Personal agents remember context and carry work between conversations. OpenClaw and Hermes helped put that model into practice in open source.
• Dots is a personal agent in ChatGPT. Muse works through its app and WhatsApp. Grok Bot supports cooperating Bots and shared Team Bots. Instinct leads with texts and calls.
• Access to an app, permission to act and an allowance for sustained work are different things. Each affects what can actually be delegated.
• Approval controls help supervise an agent. The terms still make the user responsible for its actions, and Instinct's licence specifies personal use.

What makes an agent personal?

A personal agent is an AI assistant that keeps a person's context and carries work forward across conversations. It can remember preferences, use connected tools, pick up a task later and return when it has a result or needs a decision. The relationship continues after a single request is answered.

That model needs several parts to work together. Memory carries context. Tools let the agent do something with it. A running process lets it continue when the person steps away. Permissions determine which actions it can take. A strong model helps, but the product also has to keep those other parts working.

OpenClaw made this arrangement visible well before the September launches. In his January 2026 introduction to OpenClaw, Peter Steinberger described an open agent platform that runs on a person's chosen machine and follows them through chat apps. He traced the project to November 2025. Its current documentation describes an always-on gateway for personal assistants and shared team use.

Hermes Agent, built by Nous Research, brings persistent memory, scheduled tasks and skills that it creates and improves through use. It can run on chosen infrastructure and communicate through messaging platforms. Hermes also offers a desktop app and optional cloud deployment, making that open-source model easier to run.

OpenClaw and Hermes helped pave the way by assembling these parts in products people could use and extend. Now the large AI providers are taking on more of the setup and operation themselves. That opens a path to a much larger audience, including people who want an assistant without running an agent system.

The recent releases show how quickly that path is developing. Grok Bot launched on 11 August 2026. Meta introduced Muse on 8 September. Team Bots followed on 28 September, then OpenAI dots and Muse for Small Business on 29 September. Wider distribution is the direction; access still varies by plan, country and rollout.

Dots, Muse, Grok Bot and Instinct at a glance

This comparison reflects the documentation checked on 30 September and 1 October 2026. Scroll the table horizontally to see all four products.

What differs	OpenAI dots	Meta Muse	Grok Bot	Instinct
Who it works for	Personal dot; specialist organisational dots are separate pilots	Personal Muse, with new business connectors	Cooperating Bots; Team Bots add shared roles	Personal assistant reached by text or call
Where it works	Cloud computer; optional local access	Muse Secure VM and browser	One shared cloud computer per user	Phone and computer use described; hosting details not established here
How people reach it	ChatGPT; Slack and Teams described at launch	Muse app and WhatsApp	Desktop, iOS and Android; Team Bots can join Slack	Text messages and calls
Approval controls	Custom rules, Auto-review and required handoffs	Sentinel checks actions against permissions	Auto Review and ask-first rules	No comparable approval workflow established by the checked public pages
Access	Pro, Business Premium and admin-enabled Enterprise beta	Latest announcement names US and Canada	Eligible Grok/Cursor plans; newer Teams and Enterprise beta	Sign-in path is public; current pricing and regional access not verified

The sections below explain these differences and link to the supporting sources. The products sit among the systems that hold context, organise work and reach other software in our agentic economy market map.

OpenAI dots: one personal agent inside ChatGPT

Dots gives ChatGPT users a continuing point of contact for work. OpenAI's launch describes an agent with its own cloud computer and browser that can use connected apps, pursue several projects and take feedback along the way. The starting offer is a personal dot. Teams of dots are a future direction; specialist dots for organisations have a separate preview and pilot path.

A personal agent carries someone's priorities across tasks. A company agent may need its own identity, access rights and a role that survives staff changes. OpenAI is pursuing both, but the personal rollout and specialist pilots are different offers.

The dot's cloud computer is also separate from a user's laptop. Local access starts turned off. Giving the dot access to that device is an additional choice.

Meta Muse: an assistant connected to everyday apps

Muse works through its own app and WhatsApp. Meta describes a dedicated computer, Muse Secure VM, where it can use a browser and connected applications. It remembers relevant details and keeps working on longer tasks after the app closes.

Muse for Small Business adds connections to tools including Canva, Shopify, Stripe and QuickBooks, alongside Facebook and Instagram business accounts. Meta also describes custom connectors for services outside its supported list.

Those connections give Muse more of the records it needs to understand a business. They can save someone from explaining their brand, storefront or books at the start of every task. The result still depends on the quality of those records and how well the agent uses them. Our chapter on tools, browsers and service access explains the difference between reaching software and getting useful work done through it.

Grok Bot: several Bots, or a shared team assistant

Grok Bot's launch makes cooperation between agents part of the offer. Bots can message one another, coordinate in group chats and use routines learned from a demonstrated workflow.

Team Bots adds shared roles. A team can give a Bot files, instructions, skills, plugins and credentials, then let colleagues work with that expertise. Their individual conversations, context and memories remain separate. The Bot can also join a Slack channel through its own handle.

Several Bots working for one person and a Bot shared by several people solve different problems. The first divides work between agents. The second also has to separate team knowledge from private conversations.

There is another distinction beneath the Bot names. Grok Bot's access guide says all Bots owned by one user share a cloud computer, including its files, browser sessions and command-line credentials. Naming a second Bot does not give it a separate set of account permissions. Our runtime and orchestration chapter examines how agents divide and coordinate work.

Instinct: a personal assistant you text or call

Instinct belongs in this class too. It describes an assistant connected to applications and devices, including email, messaging, screen, audio and location. Its main interface is a conversation by text or phone. The company says it can follow up on dropped threads and call or text proactively.

That puts the assistant closer to a contact in the phone than a workspace someone opens to manage agents. It is a different way to make personal delegation familiar.

Its privacy policy describes access to connected information according to the permissions granted, including private messages and, in some cases, account credentials. It also distinguishes disconnecting an integration from deleting data already collected. Those choices matter for an assistant intended to learn a great deal about its person.

We have not verified current prices or a complete list of supported countries. The public site leads to a sign-in flow; that does not establish that anyone in any country can start using it. Its terms also describe possible paid features, so we would not label the service universally free.

What can they do without asking?

The agents' ability to keep working depends on permission as well as tools. Their documentation describes different ways to handle it.

Dots: OpenAI's safety design separates proactive research from assigned tasks. Background research uses restricted read-only tools; it cannot directly send messages, change connected-app content or operate a browser. Later actions go through the usual rules. The safety FAQ describes custom rules and action review, while steps such as transferring money or changing a password require the person to take over.

Muse: Meta's Sentinel design checks actions and network requests against permissions. It supports grants with different scopes and durations. Read-only or already permitted actions can proceed without a new interruption. Purchases have a specific approval rule: Meta says browser checkout using stored payment details requires approval of the exact purchase every time, as does each payment through its Link wallet.

Grok Bot: With Auto Review on, its approval guide describes rules to ask first or allow matching actions automatically. If both rules match, ask first wins. Team rules can impose stricter controls. The guide calls for handing sensitive sign-ins and payment confirmations back to the person.

Instinct: The checked public pages do not document a comparable approval workflow. Its terms allow it to act across connected services; the scope of that authority matters, as the next section explains.

Reading an inbox, drafting a reply and sending it are three different permissions. The same applies to finding an offer, preparing an order and buying it. Our chapter on identity, authority and policy separates those steps.

What the terms say about responsibility

An approval feature helps control what an agent does. The contract describes who accepts the consequences. We read the product-specific clauses alongside the security documentation because both matter when someone delegates work.

Dots: Section 14 of OpenAI's Service Terms, updated 29 September, explicitly covers dots and agentic features. It makes the user responsible for the access and authority granted, actions taken, content shared and transactions made on their behalf. It also requires oversight appropriate to the task, including limits and review. The underlying agreement differs between personal and business accounts.

Muse: Its Supplemental Terms, dated 8 September, put responsibility for instructions, commitments and transactions on the user. They require meaningful human oversight and appropriate approval or pre-authorisation for actions that change data or the outside world. Some permissions can cover future actions, while the documented purchase flows still require approval each time. The terms also disclaim guaranteed execution.

Grok Bot: Its supplemental terms, updated 3 September, make the customer responsible for permissions, approval requirements, spending limits and unattended activity. They warn that automated review may miss unintended or harmful actions and cannot reverse completed ones. The controls assist supervision; the terms do not promise that those controls eliminate the consequences of a mistake.

Instinct: Its terms, revised 26 August, appoint the service to enter agreements and commitments on the user's behalf and describe those commitments as binding on the user. They allow possible confirmation safeguards without guaranteeing that they prevent errors. Section 6 grants use for personal purposes only, which matters when comparing Instinct with products explicitly sold for teams and businesses.

These clauses show how the vendors allocate responsibility. How they apply in a dispute depends on the governing agreement and applicable law. For the market comparison, the consequence is clear: a personal agent can carry out work while the person still carries the obligations created by it.

Who can use them, and what work is included?

“Always-on” describes availability, not an unlimited supply of work.

Dots access covers Pro, Business Premium and an Enterprise beta that needs admin enablement. Plus is absent from the launch list. The first dot is included with Pro or Business Premium, with an allowance for deeper work. Tasks it starts or manages in Codex or ChatGPT Work still count towards those products' usual limits.

Meta describes free Muse access and optional subscription plans. Grok Bot's current overview lists every paid individual Cursor plan, Cursor Teams and linked SuperGrok subscriptions, with usage resetting weekly. Team Bots has a newer public beta on Teams and Enterprise plans. Instinct's checked public pages do not give us a comparable price or work allowance.

For Australia, Muse's latest business announcement names the US and Canada. Dots has plan-specific regions: Pro excludes the EEA, Switzerland and UK; Business Premium covers supported ChatGPT regions. Its rollout is gradual, so those rules do not confirm access for a particular account. We have not verified a complete Australian eligibility statement for Grok Bot or Instinct.

A subscription price alone would be a poor cost comparison. How much work an account includes, how quickly it runs out and how often a person must step in will affect the cost of using an agent. Those questions need comparable trials, which we haven't run.

Why personal agents matter to the agentic economy

The change is that more people may have an agent that holds their brief and carries work forward. OpenClaw and Hermes showed how that could work in open products. The large providers are building routes to a wider audience, while Instinct shows that a new entrant can compete on how naturally people reach their assistant.

For businesses, that raises a new question: how does a service become useful to someone's agent? The agent needs to understand the offer, reach the provider, pass on the right instructions and check what comes back. The customer can keep the relationship with their personal agent while the agent brings in other software and services.

We see that as one route from personal assistants to an agentic economy. It remains a direction of travel, rather than a completed market demonstrated by these launches. App connections, authority to buy and payment are still separate parts. Meta, for example, describes Muse checkout through Link, built by Stripe. That is a specific payment route. Our payment credentials and settlement chapter maps the wider set.

The next useful evidence is whether these agents finish worthwhile work repeatedly, with manageable cost and supervision. The product announcements make the category easier to understand. Delivery will decide how far it goes.

Our Hitchhiker's Guide to the Agentic Economy examines the companies and infrastructure behind that change.

Agentic Economy researches how agents change the way services are found, bought and delivered. Documentation checked 30 September and 1 October 2026. About Agentic Economy.