Grok Bot, Muse from Meta, OpenAI's dots, and Claude Cowork made always-on personal agents widely available. Their vendors also offer team products, including OpenAI's Team Tasks and Anthropic's Claude Tag. This guide helps you choose the ownership and review model your work needs, and evaluate the specific product that provides it.
Use a personal AI agent for one person's research, documents, inbox, and follow-ups. Choose a team-owned operation when the company answers for the outcome: triggers from business systems, a reviewer other than the requester, continuity after the author leaves, and a shared run record. OpenAI's Team Tasks and Anthropic's Claude Tag are team options to evaluate alongside specialist platforms. Check the particular product's permissions, approval controls, and ownership; a vendor's personal assistant does not define its whole offering. Most teams will use both.
Between 11 August and 29 September 2026, three of the largest AI companies shipped the same idea. Grok Bot launched in beta on 11 August, Muse from Meta on 8 September, and OpenAI's dots on 29 September. Anthropic's Claude Cowork, already a personal coworker, now runs in Anthropic's cloud by default, so its scheduled tasks run with no device online.
Read their own documentation and the shape is consistent. Each is an agent you talk to. Each keeps working in the vendor's cloud after you close the laptop. Each can run on a schedule, and three of the four can react to events in connected apps. Each can stop and ask before a consequential step. Each keeps a record of what it did.
Those used to be selling points for automation platforms, ours included. They aren't any more, and pretending otherwise would make this guide worthless. If the question is whether an AI can do a piece of work in the background and check in before the risky part, the answer from all four is yes.
The strengths are real and large, so start there.
Drafting, research, your own follow-ups, errands, a weekly summary only you read: a personal agent is built for this, and a team operations platform is the wrong tool. Some of it you should keep manual, too. Nothing below argues otherwise.
The line isn't capability. It's ownership. Run a workflow through these seven questions. If most answers come back "one person," use a personal agent and stop here. If two or more come back "the team" or "the company," the workflow wants an owner that isn't a person.
If a customer, an auditor, or a manager would ask the company why something happened, the workflow needs team ownership. Refunds, onboarding, vendor payments, customer replies, access changes. A dot remains a personal agent even though its messaging guide now lets you instruct it to engage with others in Slack. Shared conversation and team-owned operations are separate questions; Team Tasks and Claude Tag address the latter in different ways.
Check the particular product's triggers. Personal agents can start from messages, calls, schedules and supported app events; dots also support MCP Events, and Grok Bot routines can start from Slack and GitHub events. OpenAI's separate Team Tasks run on schedules or supported events under a team service account. For a billing webhook, public form, shared email address or API call, check whether the product supports that exact entry point and whose credentials it uses.
The personal products compared here can ask before a consequential step: Muse has accept/reject cards, Grok Bot offers Allow once/Always allow/Deny, Cowork's manual mode asks for each action, and a dot sends requests you open and answer. For team products, check the actual reviewer model. If a finance lead or someone on call must correct an amount before execution, ask specifically about reviewer assignment and argument editing. Shared Slack access or a team service account alone does not establish those capabilities.
A personal agent lives in its owner's account. Grok Bot is explicit about this even for its shared Team Bots: a routine "runs as them, reports in their chat," and "no routine runs for the whole team at once." Work a team depends on should outlive one person's holiday, role change, or departure, with a definition the rest of the team can read and change.
Personal agents enforce standing rules in different ways. OpenAI calls a dot's custom rules "instructions your dot tries to follow, and it can make mistakes"; SpaceXAI's own guide says "with Grok Bot, the rules are a prompt." Meta takes the other approach with Muse: it describes a separate component, Sentinel, as the sole permission authority that decides whether an action is allowed, denied, or needs the user. Claude Cowork lets you mark each connector tool always allow, needs approval, or blocked. If "never send without approval" is a requirement rather than a preference, look for enforcement that doesn't depend on the model's judgment.
Personal agents keep records for their owner: a dot's Activity view, Muse's audit trail of everything it has done and plans to do, Grok Bot's 20 most recent run records per routine, a Cowork session's history. Some enterprise tiers add admin audit exports. When a colleague has to open last Tuesday's run and see what was decided, by whom, and with which arguments, the record needs to belong to the workflow rather than to whoever set it up.
Each personal agent settles the model question for you. A dot runs on GPT-6 Astra, and Muse on Muse Spark. Claude Cowork runs Claude models, through a Claude account or your own Amazon Bedrock, Google Cloud, or Microsoft Foundry. For Grok Bot, "Cursor manages model selection" and there's no customer-facing picker. That's fine for one person. A team with a procurement decision, a provider it already pays, or keys it must control needs the model to be a setting.
OpenAI's Team Tasks are separate team-owned automations: scheduled or event-triggered work under a team service account and configured connections, with model selection and Compliance API records for admins. Anthropic's Claude Tag is a separate public beta on Team and Enterprise: shared Slack channel work under its own service accounts, with admins configuring access. Duet is another product, a US waitlist research preview for individual Pro, Max and Ultra subscribers. Evaluate the team products for work that needs shared ownership.
Grok Bot's Team Bots, in public beta since 28 September, give a team one shared Bot that can have its own Slack app. Meta describes Muse for Small Business as "a collection of new skills and connectors inside Muse." Team ownership, shared conversation and governed tool access can take different forms; none alone establishes that a reviewer can edit an arbitrary proposed action.
If one of those fits your workflow and your team already lives in that vendor's product, evaluate it seriously before anything else. The seven questions still apply: what starts it, who approves, whether the reviewer can change the action, and whose record it is. Where a vendor's public docs don't answer one, ask them directly.
| Personal agent | Team-owned operation | |
|---|---|---|
| Who directs it | One owner, in conversation | A workspace, through a written definition and roles |
| What starts work | The owner's messages, schedules, and events in apps they connected | Business events as well: webhooks, forms, inbound email, APIs, other systems |
| Who approves | The owner | An assigned reviewer, who need not be the author |
| Changing a proposed action | Check the personal product's approval and draft-editing controls | Check reviewer routing and whether exact arguments can be edited |
| When the author leaves | The agent goes with their account | The definition stays with the team |
| The record | The owner's activity history | One record per run that colleagues can open |
| Model | The vendor's | A setting, often with your own keys |
| Best at | Conversation, browsing, personal follow-through | Recurring work the company answers for |
| OpenAI's dots | Muse from Meta | Grok Bot | Claude Cowork | |
|---|---|---|---|---|
| Who directs it | A personal account; Slack replies to you by default, with engagement with others when instructed | Its owner, on their own Muse Secure VM | Its owner; Team Bots are shared, but routines stay personal | The session owner; sessions can't be shared |
| Starts work from | Chat, calls, Slack, schedules, supported app events, MCP Events; US Pro texting beta | Chat in the app, WhatsApp, web, or Mac; schedules; relevant events | Chat and voice; routines on a schedule or after a Slack or GitHub event | Chat, or a schedule: hourly, daily, weekly, weekdays, or on demand |
| Approvals | Automatic review decides: proceed, ask you, or hand the step to you | Approval cards with accept or reject | Allow once, Always allow, or Deny, plus model-based Auto Review | Manual (allow or deny each action) or automatic approval |
| Standing rules | Custom rules it "tries to follow" | Sentinel, a separate permission authority, allows, denies, or asks | Natural-language rules; "the rules are a prompt" | Per-connector always allow, needs approval, or blocked |
| Record | Activity view; Compliance API records on Enterprise | An audit trail of what it has done and plans to do | 20 most recent runs per routine; Action Recording on Enterprise | Session history; OpenTelemetry export for organizations |
| Model | GPT-6 Astra | Muse Spark | "Cursor manages model selection" | Claude, direct or through Bedrock, Google Cloud, or Foundry |
| Team or business option | Team Tasks, a separate feature | Muse for Small Business: skills and connectors inside Muse | Team Bots, in public beta | Separate Claude Tag public beta: Team/Enterprise, shared Slack channels, admin-governed access |
Most teams will run both, and the boundary is easy to state: personal agents for the work a person directs, a team operation for the work a team answers for. Three habits keep it clean.
We make Boring AI, a console for supervised automation, so read this section as the vendor's view. Team ownership and governance are also available from other vendors. Boring's approach combines business-event triggers, assigned reviewers who can modify exact tool arguments, and a shared run record.
If what you need is a capable assistant for your own work, we're the wrong tool, and one of the four above is the right one. If a workflow answered "the team" to most of the seven questions, that's the work we're built for, and white glove means our team can build and operate it with you.
An AI agent that works for one person. It takes direction from its owner in conversation, keeps working in the vendor's cloud after they close the app, runs on schedules and some app events, asks its owner before consequential steps, and keeps a record of what it did. OpenAI's dots, Muse from Meta, Grok Bot, and Claude Cowork are the best-known examples as of October 2026.
Sometimes, but check ownership and controls for the specific product. A personal agent acting on one account is different from OpenAI's Team Tasks, which run team-owned scheduled/event work, or Anthropic's Claude Tag, a Team/Enterprise public beta for shared Slack channels with admin-governed access. The same vendor can offer both personal and team products. Evaluate those alongside specialist platforms, including their actual triggers, reviewers and access model.
Yes. The personal products here can ask before consequential steps, and Muse describes permissions enforced outside the model. For a team workflow, check who receives the request, whether another reviewer can be assigned, whether exact arguments can be changed, and what happens if nobody decides. Team Tasks and Claude Tag provide team ownership or governed shared access; those facts alone do not establish arbitrary editable approvals.
OpenAI describes a dot as a personal agent. Its messaging guide now allows engagement with others in Slack when instructed, which is different from team ownership. OpenAI's separate Team Tasks are team-owned scheduled or event-triggered tasks using a team service account and configured connections, with a model chosen per task and admin review through the Compliance API. Evaluate those when comparing team operations products.
That's a security and policy decision for your own team, and each vendor publishes admin controls to support it. The operational point is narrower: use personal agents for personal work, and move any workflow the company answers for — customer-facing, financial, access-changing — into something the team owns, so it has an owner, a reviewer, and a record that don't depend on one employee.
Boring AI is our product: a console for supervised automation a team owns. Agents belong to a workspace, start from business events as well as schedules, wait for an assigned reviewer who can approve, modify, or reject, and leave one record per run. Linked teammates can also start enabled agents in Slack and follow up in the same thread, with the same permissions and review controls. It does not offer WhatsApp conversations or computer use, so many teams will run a personal agent and Boring side by side.
Tell us the workflow you would hand off first — or ask about white glove and our team will build and run it with you.