Personal AI agents compared
No published study yet measures whether these agents save people time or money, or how often they get things wrong. Until one does, our rating for letting an agent act for you is "not enough evidence yet". This page sets out what each company says its agent does, so you can decide what to let it do. Read the guide: should you let an AI agent act for you?
For individuals, side by side
For organisations
Each agent in brief
Claude Duet
PreviewA separate app in which Claude works as a proactive agent: a morning brief, check-ins on your connected apps every few hours, and background tasks, asking before it sends, pays, signs or deletes.
- Research preview by waitlist, US Pro and Max subscribers only.
- Uses the plan's usage limits, background checks included.
Tasks in Gemini (formerly Spark)
BetaAnnounced as Gemini Spark at Google I/O and now called Task, a mode in the Gemini app: it runs scheduled and multi-step tasks across Gmail, Docs, Calendar and other apps on a remote computer and browser.
- Google still calls task handling experimental and gives no general-availability date.
- It only works with Keep Activity on, and with that on Google trains its models on your activity.
Muse
Rolling outMeta's personal AI agent, in its own app, on the web and in WhatsApp: it turns a goal into a plan, keeps working after you close the app, and books, shops and emails, with a separate "Sentinel" check on every outside action.
- Meta says more than 10 million people in the US and Canada have set it up (21 Oct 2026).
- Your conversations train Meta's models by default; you can turn it off in Settings > Data controls.
dots
Rolling outPersistent agents in ChatGPT (OpenAI writes the name in lower case): each dot works toward your goals on its own cloud computer, around the clock, through more than 4,000 connected apps.
- One dot per person at launch; OpenAI says more will come.
- Rules let you choose, per kind of action: act, act if pre-approved, ask first, or hand it back to you.
Grok Bot
AvailableAn AI teammate that signs in to your apps and websites on its own cloud computer, runs routines on a schedule, and hands back finished work. Built with Cursor, now part of SpaceX Super Intelligence (formerly xAI).
- xAI's own pages disagree on purchases: one says you approve every purchase, another that a saved card fills in without asking each time.
- There is no fixed model: each task is routed to an xAI model or a third-party one.
Gemini agent
Early accessA single agent for work in Gemini for business (Gemini Enterprise): it takes objectives, works for hours or days across Workspace, Microsoft 365, Slack and 100+ tools, under the company's rules and spending caps.
- Announced 8 Oct 2026; Google said on 16 Oct that its capabilities are not publicly available yet.
Copilot Autopilot
PreviewAn always-on agent with its own identity inside an organisation's Microsoft 365: given a role and a goal, it follows up threads, runs recurring work and coordinates meetings. Announced in June as Microsoft Scout.
- Microsoft documents a cloud version and a desktop app (Scout); most controls are documented only for the desktop app.
Before you let an agent act for you
- Start with read-only workLet it research, read and draft before you let it send, buy or delete. Every agent here can be told to ask first.
- Connect only what the task needsEach app you connect is something it can act in. Disconnecting stops new access, but what it already learned may stay until you reset its memory.
- Read the approval cardWhen it asks, check the recipient, the amount and the exact text. An approval you give "always" applies next time too.
- Keep money behind your own handUse the one-time or merchant-limited cards where offered, set a usage cap, and never give it banking logins.
- Check what it keeps and what trains the modelSome agents use your activity to train their models unless you opt out; one can't be used with that setting off.
- At work, ask before you connect work accountsA personal agent with access to work email or files is a tool your employer has not approved. Ask IT, or use a workplace version.
What changed since launch
October 2026 · 26 changes
- MusePlans and reach Launched in Singapore. source
- Gemini agentNew ability Private network connections for company tools (MCP servers). source
- Grok BotRules and safety Permissions settings; automatic review asks less often. source
- Grok BotRules and safety Bots can propose changes to their own review rules, on a card you confirm. source
- Claude DuetRules and safety Safety, privacy and security article published: approval flow, sensitive-information checks, isolated cloud browser. source
- MusePlans and reach Launched in Japan, its first market outside North America. source
- Grok BotNew ability Vault for logins, cards and contact details. source
- dotsRules and safety Help centre adds "aged 18 and older" to the plan requirements. source
- dotsProblem Service incident: degraded performance of dots for Enterprise customers (resolved). source
- Grok BotRules and safety Deleting a Bot now erases its conversations, memory and routines; admins can delete members' data. source
- Gemini agentRules and safety Admins can set usage limits per person. source
- Claude DuetLaunch Research preview opened by waitlist for US Pro and Max subscribers. source
- MusePlans and reach Meta says more than 10 million people in the US and Canada have set up Muse. source
- Gemini agentPlans and reach Government version in preview. source
- Gemini agentNew ability Claude Opus 5.5 and Sonnet 5.5 available in the Gemini Enterprise app, if an admin turns them on. source
- Tasks in Gemini (formerly Spark)New ability A chat-to-tasks toggle in the Gemini app; Google's help pages now call Spark "Task". source
- dotsPlans and reach Dots can be set up in the ChatGPT phone apps, not only on desktop (approximate date). source
- MuseNew ability Muse Wonder models announced, to power Muse "soon", subscribers first. source
- Grok BotPlans and reach Included in the new X Subscription: Premium $30, Super $100, Ultra $200 a month. source
- Grok BotNew ability Bots get their own email address; payments with X Money or Link. source
- Gemini agentPlans and reach Early access with selected small-business customers, "available to all customers soon". source
- Gemini agentLaunch Announced at Gemini at Work 2026, in early access; financial services and legal versions in preview. source
- Copilot AutopilotNew ability Dynamics 365 sales and service skills in public preview. source
- dotsProblem Two service incidents: failed and delayed dot turns, reduced proactivity (resolved). source
- Grok BotPlans and reach Tag @bot on X to give it work (not in Australia, the UK or the EU). source
- dotsProblem Service incident: elevated errors across ChatGPT, including dots (resolved the same day). source
September 2026 · 16 changes
- MuseRules and safety Launch post reworded on what Muse can see of passwords typed in its browser. source
- dotsPlans and reach Help centre rewords pricing: the first dot is included at no extra cost, with extended limits for a month. source
- MusePlans and reach Muse for Small Business, with business connectors; Meta lists the US and Canada. source
- dotsLaunch Launched at DevDay 2026 on GPT-6 Astra, for Pro and Business Premium; beta for Enterprise. source
- MusePlans and reach Meta Enterprise Platform announced, to bring Muse to businesses. source
- Grok BotNew ability Team Bots, shared in Slack, in public beta. source
- Copilot AutopilotLaunch Renamed Autopilot in the new Copilot: a cloud agent in the organisation's tenant, billed by usage. source
- MuseNew ability Connect 2026: voice mode, new shopping connectors, its own email address and glasses announced as coming; a Mac app with computer use is out. source
- Grok BotNew ability Grok 4.7 released, trained to work inside Grok Bot. source
- MuseRules and safety Muse privacy policy takes effect. source
- Grok BotNew ability Routines can be paused; GitHub connection. source
- MusePlans and reach Prices listed: free with a weekly limit, Power $20 and Max $100 a month (earliest archived copy). source
- Grok BotNew ability Can route its web traffic through your own computer. source
- MuseLaunch Launched in the US on iOS, Android, the web and WhatsApp, on Muse Spark. source
- Grok BotRules and safety Email and Slack drafts for review; admins can enforce automatic review of actions. source
- Grok BotPlans and reach Grok Bot for Enterprise opened to every Enterprise team. source
August 2026 · 7 changes
- Grok BotNew ability Works with X: an X connector and free X API credits for paid users. source
- Grok BotNew ability Purchases through Link, each with an Allow or Deny card; shareable Bot templates. source
- Grok BotPlans and reach Included with all SuperGrok, Cursor Pro and Cursor Teams plans. source
- Grok BotPlans and reach Usage limits you can set, and access for Pro and SuperGrok plans. source
- Copilot AutopilotNew ability Desktop release: meeting transcripts, room booking, Outlook rules, file sharing (August; no day given). source
- Tasks in Gemini (formerly Spark)New ability Moves to Gemini 3.7 Flash. source
- Grok BotLaunch Launched in beta on desktop and iPhone for SuperGrok and Cursor subscribers; Enterprise waitlist. source
July 2026 · 5 changes
- Tasks in Gemini (formerly Spark)Plans and reach Expanding to AI Pro subscribers in Canada and Latin America. source
- Tasks in Gemini (formerly Spark)New ability Works with Chrome auto browse, using your logged-in accounts with permission. source
- Tasks in Gemini (formerly Spark)Plans and reach AI Pro subscribers in 160+ more countries. source
- Tasks in Gemini (formerly Spark)Plans and reach Rolling out to Google AI Pro subscribers in Australia. source
- Copilot AutopilotNew ability Desktop release: model choice, Teams channel posting, session history in OneDrive (July; Microsoft gives no day). source
June 2026 · 2 changes
May 2026 · 2 changes
- Tasks in Gemini (formerly Spark)Launch Announced at Google I/O as Gemini Spark: trusted testers first, then a beta for US AI Ultra subscribers. source
- Tasks in Gemini (formerly Spark)Plans and reach Google AI Ultra gets a new $100 tier and its top tier falls from $250 to $200; Spark is listed as an Ultra benefit. source
No changes of that kind for this agent yet.
How this page works
- The company's own word, checked. Every fact comes from the company's own announcement, product page, help centre or documentation, never from news reports. Where they disagree with each other, we say so.
- No ranking. We do not pick a best agent. Which one fits depends on what you use, where you live and what you will let it do.
- Gaps stay visible. "Not stated" means the company does not say. It is a question worth asking before you sign up.
- Rechecked as they change. These products change weekly. The date at the top is the last full check; changes go in the log above.