Guide Personal Agents · Chapter 2

    What a personal agent gets done on a Monday.

    It sorts the inbox, moves appointments, researches with sources and runs the Friday routines, from a messenger on the move too, and asks before it sends.

    As of: October 2026 · Editor: Marcel Haas

    • Inbox
    • Appointments
    • Research
    • Routines
    • Messenger
    • Computer use

    Capability matrix

    Task against agent.

    Every cell says whether an agent gets a task done today, gets it done with limits, or not yet; the status per agent is in the column.

    TaskCopilot AutopilotStatus: Announced as a private preview on 25 September 2026, through the Frontier programme for tenants with a Copilot licenceChatGPT dotsStatus: Available since 29 September 2026; in Switzerland through Business Premium and as an Enterprise betaGrok BotStatus: Beta since 11 August 2026 on the SuperGrok plans, enterprises on a waitlistMuseStatus: US only so far; no dated status with a source on our pagesOpenClawStatus: Open source under the MIT licence, October 2026 per the OpenClaw Foundation's documentationHermes AgentStatus: Open source under the MIT licence; no dated status with a source on our pages
    InboxRead mail, sort by urgency, replies as drafts, file receipts
    with limitsIn Teams and Outlook with the person's access once the preview reaches the tenant; Cowork until then
    with limitsThrough the plugins and connectors the admin approves for the dot
    with limitsSigns in with stored access; test accounts only in the beta
    with limitsWith stored credentials on the machine at Meta; no source on our pages
    todayThrough Microsoft Graph with delegated permissions; drafts in Outlook, sending only after approval
    with limitsThrough an MCP server for mail and calendar that you run
    AppointmentsFind free slots, move, confirm, prepare papers, record commitments
    with limitsWith the person's access in the tenant; preview
    with limitsThrough approved connectors; check-back before every commitment
    with limitsMulti-step work up to the approval; test accounts in the beta
    with limitsIn the connected services; no source on our pages
    todayPropose and move through Microsoft Graph; commitments to outsiders wait for the yes
    with limitsThrough an MCP server that you run
    ResearchWatch competitors, tenders and law changes, with sources
    with limitsWith its own computer and web search; preview
    todayKeeps working towards goals on its own cloud computer, even when nobody is in the chat
    todayWeb and X from its own cloud computer, with linked sources
    with limitsNo source on our pages
    todayWith the model of your choice; web search can be switched off where it is not needed
    with limitsWith the model of your choice; bounded toolset
    RoutinesWeekly report, expenses, time sheet, follow-ups, every Friday on a schedule
    with limitsKeeps working persistently; schedules per task we check in the preview
    with limitsGoals keep running; fixed schedules per dot we check in the workspace
    with limitsMulti-step work up to the approval; schedules we check in the beta
    with limitsNo source on our pages
    todayHeartbeat and fixed schedules through the gateway, a skill per routine
    todayDerives skills from completed tasks; schedules on your machine
    On the move by messengerStart a task from the phone and receive the result there
    with limitsIn Teams on the phone; not over WhatsApp, Signal or Telegram
    todayIn ChatGPT, by voice, in Slack or in Teams
    with limitsWhere Grok runs; not over WhatsApp, Signal or Telegram
    with limitsIn the Muse app and in the browser; no source on our pages
    todayGateway for Teams, Slack, Telegram, WhatsApp and Signal, with a pairing code for unknown senders
    with limitsDepending on the setup on your machine
    Computer useOperate a program without an interface through the screen
    with limitsOwn computer and workspace in the tenant; preview, with test data
    with limitsOn its own cloud computer; which applications, the admin sets per dot
    with limitsSigns into tools from its cloud computer; beta, with test accounts
    with limitsVirtual machine at Meta; no source on our pages
    with limitsTools in the sandbox with Docker or Podman; screen control only with approval and test data
    with limitsIn the container, with a bounded toolset
    Your own knowledgeYour own notes, quotes and contracts as memory, carrying on in the same tone
    with limitsMemory in preview in the person's mailbox; documents with the person's rights
    with limitsDocuments from approved connectors; memory in the workspace at the vendor
    with limitsDocuments you give it; memory on the cloud computer at the vendor
    with limitsData on the machine at Meta; no source on our pages
    todayMEMORY.md and USER.md in your storage, files through MCP servers
    todayRemembers across sessions; MEMORY.md in your storage
    Legend:todaywith limitsnot yet

    As of October 2026, derived from the dated facts of this chapter, the vendor pages that back them and the anatomy of the personal agent. Muse and Hermes Agent carry no dated status with a source; their cells follow the status of the Personal Agents page as of 3 October 2026. Availability and plans change; before the start we check with the vendor what has changed since.

    01

    Monday, 06:45: the inbox is sorted before you open it.

    The head of sales of a machine builder opens the laptop on Monday and finds the mail since Friday evening in three piles: at the top the four that need a decision from her, below them the routine questions with a reply as a draft in Outlook, at the end invoices and receipts that already sit in the right folder. The agent has sent nothing. It read with her delegated permissions, sorted by urgency and put the overview into her Teams chat at 06:45, where she reads it on the phone over the first coffee. Copilot Autopilot keeps working like this in Teams and Outlook while the person does other things; a dot reports with the result in ChatGPT, in Slack or in Teams; a self-hosted agent puts the overview through its gateway into the channel IT has approved. Who asks before sending is in the directive, and the result on Monday is the same: the mails that count are at the top.

    02

    Moving appointments, preparing papers, recording commitments.

    Huber AG wants to move Thursday's meeting to the afternoon. The agent's memory says the head of sales meets Huber in the morning only; it proposes a morning slot, saves the reply as a draft and waits for her yes before it uses the permission to send. For the ten o'clock client meeting it fetches the latest quote from SharePoint, the notes from the CRM and the minutes of the last meeting from OneDrive and turns them into one page with the open questions, attached to the appointment. After the meeting it records what she committed to, and tomorrow that still holds. Finding free slots, moving and confirming it may do with her consent; a commitment to a client leaves the mailbox only after a yes.

    03

    Research with sources and the Friday routines run without a prompt.

    Every week the agent watches three competitors, two tender platforms and the changes to one law, summarises them and puts the sources next to it, so the person can check every statement; the briefing is ready on Monday morning. Grok Bot plays to its strength here, because it searches the web and X on a cloud computer of its own and completes multi-step work up to the approval. On Friday at 16:00 a schedule starts the weekly report from the person's own notes, the expense report from the week's receipts, the time sheet and the follow-ups on open quotes, each as a draft the person reviews. A rhythm wakes the agent at a fixed interval, and it stays quiet when nothing is due. Chores get done before they pile up, and nobody has to remember them.

    04

    From the phone, while the laptop stays shut.

    Tuesday evening on the train: the head of sales writes to her agent to prepare Thursday's supplier meeting. Whether that message arrives in Teams, in ChatGPT, in the Grok app or over Signal depends on the agent; it keeps working on its computer, in the tenant, on the vendor's cloud computer or on the server in the office, and reports back with the result. An OpenClaw gateway serves Microsoft Teams, Slack, Telegram, WhatsApp and Signal at the same time, and unknown senders first receive a pairing code. A dot takes goals by voice as well. Whoever starts tasks from the phone needs a channel IT has approved and an agent that listens there only to the accounts on its list; messages from other senders it reads as data.

    05

    Your own notes and quotes become its memory.

    Every note, quote and contract a person has ever written can serve the agent as memory: it finds the passage, quotes it and carries on in the same tone, and the next quote starts from the best one. Added to that are the agreements and decisions it records itself, as short dated entries. Where this memory lives sets the agents apart: OpenClaw and Hermes Agent write MEMORY.md and USER.md as files into a workspace that we put in your own storage; Microsoft Copilot stores its memory, in preview, in a hidden folder of the Exchange mailbox; a dot keeps goals and history in the workspace at the vendor. Which documents it may read is decided in chapter 5, what the memory is made of in chapter 3.

    06

    What no agent does well today.

    Judgement. Whether a customer's complaint deserves a credit note, whether the supplier caused the delay, whether an application makes the shortlist: the agent prepares the papers, the person decides. Anything with consequences without a check: sending, committing, booking, paying and deleting remain tasks it prepares and then waits for a yes, and purchases above the limit in the directive it declines. Instructions that sit inside a mail or on a web page: the OpenClaw documentation states that prompt injection is not solved and that smaller models are easier to steer with manipulated content; so an agent reads other people's mail as data, in a sandbox and with the strongest model. Figures that have to be right, such as expenses and invoice amounts, a person checks before they are booked. And long sequences in programs without an interface, which it operates through the screen, break off more often than mail and calendar.

    07

    Computer use and voice are the new abilities.

    Two abilities arrived this year. An agent with a computer of its own operates programs through the screen like a person, wherever there is no interface: a supplier's portal, a booking system, an old line-of-business program. Autopilot brings a computer and workspace of its own in the tenant, a dot and Grok Bot work on a cloud computer at the vendor, OpenClaw runs tools in a sandbox with Docker or Podman. And voice: a person dictates a task to the agent on the move, and it confirms in words what it understood. We trial both with test accounts and test data, because an agent at the screen has the same rights as the person who gave it the access. Where each agent stands today, in one sentence per vendor:

    • At Microsoft: Copilot Autopilot reads, sorts and prepares in Teams and Outlook and keeps working with its own computer while you do other things; as a private preview, and until then Cowork takes described tasks from start to finish.
    • At OpenAI: a dot keeps working towards goals on a cloud computer of its own and takes them by chat, by voice, in Slack or in Teams; it reaches applications through the plugins and connectors the admin approves.
    • At SpaceXAI: Grok Bot signs into your tools from its cloud computer, completes multi-step work and speaks up when an approval is needed; as a beta, with bounded test accounts.
    • At Meta: Muse works with stored credentials on a virtual machine in the connected services, in the app and in the browser; US only so far.
    • Self-hosted: OpenClaw reads mail and calendar through Microsoft Graph, reaches files and line-of-business systems through MCP servers, runs on a schedule and reports over Teams, WhatsApp, Signal or Telegram; Hermes Agent derives skills from completed tasks and remembers across sessions.

    What you decide.

    Which task does the agent start with?

    Options

    1. 1The weekly report or the briefing from your own notes and public sources
    2. 2The inbox, reading and sorting only at first
    3. 3Appointments and replies as drafts, with an approval before every commitment

    Our recommendation: With a task without third parties' personal data that recurs every week and whose result can be checked: the weekly report from your own notes or the competitor briefing with sources. After that the agent reads the inbox, reading and sorting only at first, and only with the directive does it save replies as drafts. That way the pilot shows after a few weeks what the agent does reliably and where it gets on people's nerves, before it prepares a commitment to a customer.

    What we advise against: A pilot that starts with sending, booking or paying because that saves the most time. The approval before actions with consequences comes before the first run, and the person first learns to check results.

    What may it do without checking back?

    Options

    1. 1Read, sort, prepare and research, as much as it likes
    2. 2Plus create drafts and propose appointments
    3. 3Send, commit, book, pay and delete only after a yes

    Our recommendation: All three levels, as a rule per tool in the configuration and not only as a sentence in the instructions: reading, sorting, preparing and researching freely; drafts and appointment proposals freely, as long as nothing leaves the mailbox; sending, committing, booking, paying and deleting only after a yes in Teams or on the phone. Purchases above the limit in the directive the agent declines, and if a check-back reaches no one, the answer is no. Every approval is logged with its time.

    What we advise against: A blanket release for everything that sits in the calendar or on a list. The line applies per tool, and an action with consequences needs a person.

    How does the agent reach you on the move?

    Options

    1. 1In Teams and Outlook, where the person works anyway
    2. 2In ChatGPT, by voice, in Slack or in Teams
    3. 3Over WhatsApp, Signal or Telegram with a pairing code through a gateway of your own

    Our recommendation: Through the channel your IT has approved and on which the agent listens only to the accounts on its list. For most people that is Teams on the phone, where Autopilot, a dot and a self-hosted agent all report. WhatsApp, Signal or Telegram come on top when a person has to be reachable there, through a gateway that turns unknown senders away with a pairing code; the company agent and the private agent stay separate profiles.

    What we advise against: An agent that reads company mail through a person's private messenger account or takes instructions from any sender.

    When is computer use worth it?

    Options

    1. 1Only where no interface and no MCP server exists
    2. 2For every task, because the agent then operates any program
    3. 3Not for now

    Our recommendation: Only where there is no interface and no MCP server, such as a supplier's portal or an old line-of-business program, and first with test accounts and test data in a sandbox. Mail, calendar, files and line-of-business systems the agent reaches more reliably through Microsoft Graph, connectors or MCP servers, with exactly the permissions the task needs and a log per call. What breaks off at the screen is shown by the pilot.

    What we advise against: Computer use with the person's main account in a production system before the workflow has run with test data.

    Status and sources.

    As of: October 2026

    • Since 25 September 2026

      Copilot gets Autopilot.

      Microsoft introduced the new Copilot with Home, Code and Autopilot: a persistent personal agent with its own identity, memory and computer, announced as a private preview for the end of September.

      Source: Microsoft: Introducing the new Copilot
    • Since 29 September 2026

      OpenAI brings dots.

      Dots keep working towards goals in ChatGPT on their own cloud computer and are reachable by chat, voice, Slack or Teams. In Switzerland through Business Premium and as an Enterprise beta, not for Pro accounts at launch.

      Source: OpenAI Help Center: Getting started with your dot
    • Since 11 August 2026

      SpaceXAI launches Grok Bot.

      Agents with their own cloud computer that sign into your tools, complete multi-step work and report back when an approval is needed. As a beta on the SuperGrok plans, enterprises on a waitlist.

      Source: SpaceXAI: Introducing Grok Bot
    • Since September 2026

      Copilot Cowork is generally available.

      Cowork takes a described task from start to finish: it gathers context from emails, meetings, chats and files, works across several tools and delivers a finished result. Who reviews the result before it is passed on is set by your directive.

      Source: Microsoft Learn: Release notes for Microsoft Copilot
    • As of October 2026

      Prompt injection is not solved.

      The documentation states that smaller models are more susceptible to tool misuse and hijacked instructions. For agents with tools it recommends the strongest current model, a sandbox for sensitive tools and web search switched off where it is not needed.

      Source: OpenClaw Docs: Prompt injection

    Frequently asked questions.

    Does it read my mail while I am on holiday?

    Yes, if the rhythm and the schedules are set that way; the agent keeps working on its computer whether you open the laptop or not. A holiday rule in the directive makes sense: it sorts and prepares, saves replies as drafts and reports only what needs a decision to your deputy in Teams. Commitments, bookings and payments wait until you or the deputy give a yes.

    Which agent gets the most of this done today?

    The capability matrix above shows it, as of October 2026: a self-hosted agent with OpenClaw covers inbox, appointments, research, routines, messengers and your own knowledge today, because Microsoft Graph, MCP servers, schedules and the gateway are there for it. Autopilot brings the same in the tenant once the preview has reached your tenant; a dot and Grok Bot work on the vendor's cloud computers and reach applications through connectors or sign-ins of their own. Which one fits you is settled by the agent finder in chapter 4.

    Can the agent speak or make calls for me?

    Speak, yes: a dot takes goals by voice, and a self-hosted agent can process a dictated message from a messenger and confirm in words what it understood. Phoning on a person's behalf we do not set up; a call is an action with consequences, which the agent prepares by laying out meeting notes, open points and the last mail, and the person makes.

    What happens when a mail contains an instruction for the agent?

    It reads it as data. A mail that tells it to send the price list to an outside account triggers nothing, because sending needs an approval and the rules in its instructions keep company data in the company. The problem is not fully solved: the OpenClaw documentation states that smaller models are easier to steer and recommends the strongest model, a sandbox for sensitive tools and web search switched off where it is not needed. The agent reports such mails to you, and the log shows what it did with them.

    Does it also handle tasks in programs without an interface?

    With computer use, yes, within limits: an agent with a computer of its own operates a portal or an old line-of-business program through the screen like a person. That breaks off more often than mail and calendar, and the agent has the same rights there as the account it signs in with. So we start with test accounts and test data in a sandbox and, for systems that are used often, rather build an MCP server with exactly the tools the task needs.

    How does it remember how I write quotes?

    From your earlier quotes, which it reads as memory, and from the decisions it records: that quotes go out as PDF, that special discounts are cleared with management, that one client wants morning appointments only. With OpenClaw and Hermes Agent these entries sit as files in your own storage and move with you when you change agents; with the vendors' agents the memory lives with them, and chapter 5 says who can read it.

    How do I notice that it does not get a task done reliably?

    In the pilot with one task and one pilot group: every person records what the agent does reliably, what it gets wrong and where it gets on their nerves, such as mail it ranks wrongly or drafts nobody would send like that. After a few weeks it is settled which tasks stay, which are narrowed and whether the agent goes to more people. The directive and the rules per tool are updated with these findings.

    Next chapter · Chapter 3

    What your agent is made of, from SOUL.md to approval.

    What is a personal agent made of, and how does it decide?