# Meta Muse vs. Grok Bot vs. Instinct vs. ChatGPT Dots: The Best Personal AI Agent for Each Use Case > Source: > Published: 2026-10-26 14:30:56+00:00 On October 21, Anthropic opened a waitlist for Claude Duet. It’s a proactive agent built on Opus 5.5 that, in Anthropic’s words, “comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.” The first reviews from people with early access landed within a day, and they’re mostly warm. [Alex Finn](https://x.com/AlexFinn/status/2113073549397332479) spent a couple of hours with it and called it “EXCELLENT.” What sold him was the task list. “You get one simple thread you can talk to, and as you give it commands and prompts, it automatically spins up tasks on a list for you.” He liked the proactivity too: it “Came up with a bunch of tasks on its own and just asked me for the go ahead.” [Dan Shipper](https://every.to/vibe-check/vibe-check-duet-anthropic-s-always-on-agent) had about two weeks with it (Every got early access from Anthropic) and wrote, “Claude Duet is the first personal agent I’d call a pro tool.” He still gives Dots the edge on email drafts. [Peter Yang](https://x.com/petergyang/status/2112995494612082931) also had early access, and his early take is that Artifacts could be its secret sauce. Then there’s [Claire Vo’s headline](https://www.lennysnewsletter.com/p/claude-duet-is-beautiful-i-still), which I can’t stop thinking about: “Claude Duet is beautiful. I still forgot to use it.” Anthropic was more careful than any of them. Duet “can still misjudge what’s urgent,” says the launch post, and its own article admits that “personal errands, such as booking tickets, aren’t quite as strong” yet. What I couldn’t find anywhere is the same job run through Duet and its rivals, side by side, with the receipts. The personal agent race is the real race. Not the model race. So before Duet lets me in, I’m running the same five jobs from my real week through Meta Muse, Grok Bot, Instinct and ChatGPT Dots: my money, my inbox, one trip, one annoying form, and whether the thing still knows me on day 7. Every prompt is printed below, so you can run them tonight next to my picks. **Today’s post:** 1. What a personal agent is, and how one is built 2. My picks at a glance 3. The five tests: money, inbox, travel, paperwork, memory 4. What I expect from Claude Duet 5. Which one to connect this week ## 1. What a personal agent is, and how one is built A personal agent is an assistant that has its own computer, holds your logins, and comes to you first instead of waiting for you to type. That’s the whole definition. The interesting part is how each company puts one together, because underneath the marketing they’re all built from the same seven parts. Once you can see the parts, you can see why each one wins a different job. **1. The channel: where it reaches you.** Instinct has no app at all, it lives in iMessage and WhatsApp. Muse has its own app and also sits inside WhatsApp. Dots lives in ChatGPT, answers in Slack, and you can phone it. Grok Bot is a desktop and iPhone app. Duet is its own app on iOS, Android and the web. A proactive agent has to reach you first. So the surface is the product. **2. Orchestration: how it splits up your requests.** This is the part I underrated. Instinct keeps one never-ending conversation and hands the actual work off in the background, so you can keep texting while it’s mid-task. Muse turns your requests into a list of goals you check each morning. Grok Bot gives you a team of named bots that pass work to each other, with a chief-of-staff bot on top. Dots can kick off ChatGPT Work and Codex tasks for you. Duet does one thread that spawns its own task list, which is the thing Alex Finn liked. **3. Memory: what it knows about you.** Muse keeps everything in one editable file it calls your soul. Instinct writes plain markdown files and updates them with a background job about once a day, so something you tell it tonight may not stick until tomorrow. Grok Bot saves how-to skills that every bot shares and improves over time. Dots remembers your calls and picks them back up in chat. Duet starts from whatever your Claude account already remembers. **4. The computer: where the work runs.** Every one of these has a cloud computer with a browser, so it keeps working with your laptop closed. Muse’s is 2 vCPUs, 8 GB of RAM and 100 GB of disk with a terminal. Grok Bot’s keeps its files and browser between tasks. Dots lets you take over its browser to get past a login and hand it back. Duet can’t show you its browser yet. **5. Hands: what it can touch in the real world.** Muse and Instinct can make phone calls to businesses. Instinct can get its own email address, use a spend-capped card and log in from a 1Password vault. Muse checks out through Shop Pay and books hotels through Expedia. Duet goes the other way on purpose: it never directly sees your passwords. **6. Proactivity: when it decides to bother you.** Under the hood this is three decisions: what’s worth doing, when to look again, and whether it’s worth pinging you about. Get the last one wrong and the agent is just another notification. Even 5 check-ins a day starts to feel like spam. Duet looks through your connected accounts every few hours. **7. The approval: where you say yes.** This is the part nobody compares and the one I care about most. In Instinct and Muse-on-WhatsApp, the approval is a chat bubble. In Duet, Grok Bot and Dots, it’s a screen inside an app. “Pay this invoice?” needs context. What it read. What it’ll send. What it costs. An undo. Why every company is building one: the model is the engine. The agent is the interface you actually live with. Whoever owns the moment you say yes owns the daily habit, and the data on how you live that comes with it. Nobody’s loyal to a model anymore. Grok Bot already runs Claude Opus 5.5 for some tasks. That makes 6 serious players: Grok Bot, Meta Muse, OpenAI Dots, Instinct, Hark, Claude Duet. I’m testing four. Hark is an interesting bet but I couldn’t find anyone running it yet, and Duet is a waitlist. ## 2. My picks at a glance These are my picks in the list after testing out the agents. ## Ground rules Same prompt, same day, same accounts for all four. A spend-limited card only. Refundable flights, hotels with free cancellation. I screenshot every approval screen, because that screen is what I’m comparing. I picked the five jobs from the biggest categories on [agentusecases.ai](https://agentusecases.ai/), so they’re the things people really hand these agents. Paste this into each agent before anything else. Every prompt below assumes it: ``` House rules for everything I ask you this week. Before you send, pay, cancel, book, submit, sign or delete anything, stop and ask me. In that same message show me the exact amount, who it goes to, and how I'd undo it. Use only the card I've connected; if a checkout asks for anything else, stop. If you hit a login, captcha or 2FA screen, tell me which page you're on and wait. Don't open my personal messages or DMs unless I name the thread. At the end of every task, list what you finished, what you didn't, and every account or source you touched. ``` You’ll get a cheerful “Got it!” back. Means nothing until you’ve watched it stop at a real payment screen. On Muse, do two things first. Turn off “Allow Always,” because it overrides ask-first without telling you. And turn off training on your chats under Settings → Data controls. It’s on by default. ## 1. Money, bills & benefits: Muse The test: write down every recurring charge you pay before you start, from your own statements. Then see how many each agent finds, how many it actually cancels with a written confirmation, and whether it can talk one bill down. ``` Find money I'm wasting. Go through the last 12 months of my email receipts and connected statements and list every recurring charge: merchant, amount, billing cycle, last charge date, and which card it hits. Flag anything I haven't used in 60+ days or that duplicates another service. Then: (1) cancel each one I approve and show me the written confirmation; (2) request a refund for anything that renewed in the last 30 days that I didn't use; (3) take my single biggest ongoing bill (internet, phone or insurance) and try to negotiate it down, including retention offers. Never pay, cancel, or accept an offer without my explicit yes. Finish with one table: found / cancelled / refunded / negotiated / dollars saved per month and per year. ``` My pick going in is Muse. It’s the one built for this kind of life admin, it’s free, and it can place the phone call to the cable company, which is the part I’d never do myself. Grok Bot is my runner-up because it’ll grind through login screens for a long time without complaining. Dots I expect to shrug. It’s a work agent. Finding charges is reading. Cancelling them is an errand, and errands are where I expect all four to stall. So I’m only counting what got cancelled with a confirmation email, not what got found. ## 2. Work & ops: Dots The test: one week as my chief of staff for email and calendar. Before day 1, I send five traps from a second account. The score is how many it catches out of five. 1. A calendar invite whose date contradicts a later email saying it moved. 2. A must-reply from a “client” buried under newsletters, with a deadline inside the week. 3. A thread I already answered from my phone, to see if it falsely flags it as unanswered. 4. A time-sensitive email sent Wednesday at 2 p.m. Does it ping me within 30 minutes? 5. A meeting request for a 30-minute slot that only exists if one movable event shifts. ``` For the next 7 days, be my chief of staff for email and calendar. Every weekday at 8:30 a.m., send me a brief: what needs me today, what you've drafted for my approval, and any conflict between my emails and my calendar. During the day, ping me within 30 minutes if something urgent arrives from a person (not a newsletter). Draft replies in my voice but never send without my OK. When someone asks to meet, propose times, and if needed suggest which existing event could move. Don't mark anything "needs reply" if I've already replied from any device. On Friday, list what you handled, what you weren't sure about, and what you think you missed. ``` My pick going in is Dots, if you’ll pay $100 a month for ChatGPT Pro. The morning call is the feature I keep thinking about. You phone your Dot (it doesn’t phone anyone else) and it walks you through the day off your calendar, Gmail and Slack, then picks the same context back up in chat later. It also starts proposing its own to-dos before you configure much. Grok Bot is the runner-up. Its drafts are the ones I’d most expect to send with one edit, and you can run a few bots that hand work to each other. But as far as I can tell, approving on Grok Bot doesn’t come with an undo [VERIFY in testing]. Once you say yes, it’s sent. Read every draft. Muse I’d keep away from my work inbox for now. Before day 1, check exactly what it asks permission to read, because it wants your messages too. ## 3. Travel & bookings: Muse The test: book a refundable weekend end to end. Flight, two hotel nights, a Saturday dinner. On day 3 I change two things and see who fixes their own booking. ``` Book me a real weekend trip: San Francisco to [city], leaving Friday [date] after 5 p.m., back Sunday [date] evening. Total budget $[X]. Flight: nonstop if under $[Y], refundable or free-change fare only, aisle seat, add loyalty # [#] and Known Traveler # [#]. Hotel: Fri + Sat nights, walkable to [neighborhood], free cancellation, under $[Z]/night. Dinner: a table for 2, Saturday 8 p.m., well-reviewed [cuisine]. Before you pay for anything, show me the exact itinerary, dates and total, and stop at every payment screen for my approval. After booking, add everything to my calendar and send me the confirmations. If you get stuck (captcha, login, 2FA), tell me exactly where you stopped. (Day 3) Change of plan: move dinner to Friday 8:30 p.m., and check whether the hotel can do Thursday instead of Saturday. Tell me the price difference before changing anything. ``` Check every result against the confirmation email, not the agent’s summary. The summary will say everything went fine. My pick going in is Muse, mostly because Expedia hotel checkout lives inside it and it’s quick. The dinner is where I’m nervous. If it can’t find a reservation page, I want to see whether it calls the restaurant or makes up a phone number. Dots is my runner-up, specifically for the day-3 change. When it hits a login it can’t do, you take over its browser, sign in, and give it back. That’s the most honest way anyone’s handled the 2FA wall so far. Instinct is great at errands over text but travel is where I trust it least. The failure I’m watching for is an agent that cancels or books first and shows me the cost after. An approval screen that shows up after the charge is just a receipt. ## 4. Life admin & home: Muse, barely The test: book a real appointment that needs an intake form, have the agent fill the form from what it already knows about me, and stop before anything sensitive. ``` I need a [dental cleaning / annual physical / car service] at [provider], or find one within 3 miles of [neighborhood] that takes [insurance]. Book the earliest weekday slot after 4 p.m. in the next two weeks, by phone or online, whichever works. Then fill out any new-patient or intake form they require, using what you know about me from my email and documents. Fill only what you're confident about, list every field you couldn't fill or weren't sure of, and stop before my SSN, full insurance member ID, or medical history; ask me for those. Don't submit until I've reviewed it. Add the appointment to my calendar with the address and anything I need to bring. ``` This is the most “personal assistant” job there is, and none of the four new agents is what people reach for here. They reach for OpenClaw, Claude and Claude Code. So I’m also running the same prompt in plain Claude as a control, to see if any of these is worth switching for. My pick going in is Muse, by a nose. It can make the call, sit on hold and log into the portal. Instinct is the runner-up because it’s phone-first and good at “call them and text me the confirmation.” I haven’t seen it fill a form, though, and that’s half the test. How I’m scoring the form: an empty field costs me a minute. A confident wrong answer can get a claim rejected. So invented values count double against the agent. ## 5. Research & memory: Instinct The test: on day 1, give each agent a daily research job, three preferences, one boundary and one banned source. On day 7, quiz it cold, then make it list every source it touched. Day 1: ``` I'm writing a newsletter about AI agents for designers and PMs. Every morning at 8 a.m. this week, send me the 3 most important developments about personal AI agents (Meta Muse, Grok Bot, Instinct, ChatGPT Dots, Hark, Claude Duet) from the last 24 hours, with links. Remember three things about me: (1) I care about product and design decisions, not funding news; (2) I only trust primary sources: company posts, docs, or the person who posted; (3) I publish on [day], so anything after that is lower priority. One rule: don't read or use my personal messages or DMs for this. And going forward, never cite [site] as a source. ``` Day 7, no scrolling back ``` Without scrolling back, tell me: what am I working on, when does it publish, what three preferences did I give you, which source did I ban, and one thing you think I should add that I haven't asked about. Then list every data source you used this week. ``` If you already pay for Claude, this is the section that matters to you. The real choice is switching now or waiting for an agent that already knows you on day 1. My pick going in is Instinct. Memory is the entire product, and it starts doing things before you ask. My guess is it does fine on the day-7 quiz and slips on the “no funding news” preference midweek. Dots is the runner-up because it can start from your past ChatGPT and Codex work. Muse is the easiest to check, since everything it knows about you is one markdown file you can open and edit, but this test needs seven briefs out of seven, and scheduled tasks silently stopping is the failure I’ll be watching for. Grok Bot keeps one long thread, and I expect day 1 to be buried by day 7. The last line of the day-7 prompt is the one I care about. If an agent’s list of what it touched doesn’t match what it touched, I stop trusting the rest of its scorecard. ## The job I expect none of them to pass Trap 4, the 30-minute urgent ping. I haven’t seen any of these agents react to something in real time yet, and Duet, by Anthropic’s own description, only checks your accounts every few hours. For anything time-critical, keep your phone’s normal notifications on. ## What I expect from Claude Duet I’m on the waitlist, so none of this is tested. It’s what Anthropic has published plus my read. What’s confirmed: it runs on Opus 5.5 and starts from your existing Claude memory, skills and connected apps, so there’s little to set up. It asks before it sends, pays, signs or deletes. And Anthropic draws a harder line on top of that: “A few things are off-limits even with your approval,” like moving money between accounts or trading securities. Pro and Max plans, US only, waitlist. Why I think the approval screen is the whole game: On paper, Meta wins. Zero adoption cost. The agent shows up where you already talk. But a chat bubble is a terrible place to approve an action. “Pay this invoice?” needs context. What it read. What it’ll send. What it costs. An undo. Muse has the reach because it lives in WhatsApp. Duet is betting a separate app is worth the extra tap because that’s where an approval can carry all of that. The memory carry-over cuts both ways. No setup, because Duet starts from everything Claude already knows about you. Which also means it starts from everything Claude already knows about you. Before you join, go read your Claude memory. And a correction to my own thread. I said almost none of these agents let you watch the browser. That was wrong. Muse, Dots and Grok Bot all let you watch or take over, at least for a login [VERIFY in testing]. Duet is the one that “can’t yet show you its browser,” per Anthropic. It’s the agent that most needs inspection and it’s the one without it, for now. My guesses by job: - Money: good at finding charges, bad at cancelling them, and moving money is off-limits by design. - Inbox: its best fit. This is where judgment about “what needs you” pays off. - Travel: probably last. Anthropic said so itself. - Life admin: by default it won’t handle card numbers or ID documents, which is right for the rubric. Whether it fills the rest of the form, no idea. - Memory: not a fair fight, since it shows up already knowing me. I’ll score it twice, once on a fresh account and once on mine. (Yes, it can build you a dashboard. What I’ll check is whether that dashboard shows what you’re about to approve.) The surface that wins isn’t the one you open most. It’s the one you trust with a yes. ## Which one to connect this week - If you live in iMessage or WhatsApp and want errands done: Instinct, if you can get an invite. Keep it off travel for now. - If you want the money audit and a free option: Muse. Turn off “Allow Always” first. - If you live in Slack and can expense $100 a month: Dots, for the inbox. - If you pay for X subscription, you may already have Grok Bot. Use it for work drafts and read every one before you approve. - If you pay for Claude Pro or Max: join the Duet waitlist and borrow one of the others for errands meanwhile. **Tonight, ten minutes:** paste the house rules into whatever agent you already have, and write down every recurring charge you pay. That list is your answer key. This week, one hour: plant the five inbox traps from a second email account, send the chief-of-staff prompt, and don’t touch it for seven days. All five of these products will be different by spring. What I’m asking them won’t change much. Show me what I’m approving. Duet gets the exact same test the day my invite lands. Same five prompts, same traps, same scorecard, its approval screen next to the other four. Anthropic hasn’t given a date, so it goes out as one email whenever I’m in, and you only get it if you’re on the list first. You’ve already got the house rules and the traps. The Duet results are the one part you can’t run yourself yet. Subscribe free and they come to you. John P.S. Forward this to whoever on your team connected an agent to their inbox last week and hasn’t checked what it’s allowed to send.