Source: The AI Daily Brief (Nathaniel Whittemore): raw/How_to_Choose_Your_Personal_Agent.md, raw/AI_Agents_Are_Moving_Into_the_Real_World.md, raw/How_Muse_Wonder_Shifts_the_AI_Landscape.md, raw/Gemini_4_Argon_Sonnet_5.5_and_What_Models_You_Should_Be_Using_Right_Now.md, raw/America_s_Open_Model_Comeback_Is_Starting.md, raw/Consumers_Will_Never_Pay_for_AI.md, raw/The_Best_New_AI_Products_and_the_Key_Trends_They_Represent.md, raw/The_Biggest_AI_Announcements_from_OpenAI_Dev_Day.md · How I AI (Claire Vo): raw/Meta_Muse_-_the_AI_agent_people_actually_love.md (guest Alex Cornell, Muse design director), raw/I_tested_Claude_Duet._Is_it_worth_switching_from_Grok_Bot.md, raw/OpenAI_DevDay_2026_-_Dots_Spaces_and_ULTRAFAST.md · Lenny’s Podcast: raw/OpenAI_s_Head_of_ChatGPT_-_We_re_entering_a_new_era_of_AI_again_Tibo_Sottiaux.md (Tibo Sottiaux, OpenAI, first-party on Dots) · raw/Anthropic_Just_Dropped_Claude_Motion_Dashboards_+_More_AI_News.md (Grok Bot multi-model announcement, read word for word) · Brock Mesarich: raw/Anthropic_Just_Dropped_Claude_Duet_Full_Breakdown.md (https://youtu.be/nv1aaP3dHzs, 2026-10-21), raw/Claude_Duet_is_the_Grok_Bot_Killer_LIVE_TEST.md (https://youtu.be/PkrjD2IOCm4, 2026-10-22) · Sharbel: raw/OpenAI_Just_Launched_Dots._Here_s_The_Truth_About_It..md (Dots hands-on, added 2026-10-30) · Nate B Jones: raw/You_ve_Never_Lost_19_Billion._So_Why_Are_You_Playing_It_Safer_Than_Zuckerberg.md (https://www.youtube.com/watch?v=JFoDdKuRmSk, Muse team and scale, added 2026-11-09)
Creators: Nathaniel Whittemore, Claire Vo, Lenny Rachitsky · Platform: YouTube (auto-captions, fetched 2026-10-21) · Key URLs: https://www.youtube.com/watch?v=WloeFbIX7PU (how to choose), https://www.youtube.com/watch?v=4TOLm1g-gCE (Muse design), https://www.youtube.com/watch?v=HYb5w4PfPZM (Duet review), https://www.youtube.com/watch?v=pJNM1z9l5mU (Dots at DevDay), https://www.youtube.com/watch?v=MM-C3JqCXBk (Tibo Sottiaux)
By October 2026 there are at least nine personal agents a non-developer could pick: Meta’s Muse, OpenAI’s Dots, SpaceX AI’s Grok Bot, Anthropic’s Claude Duet, Nous Research’s Hermes, OpenClaw, Instinct, Poke and Google’s Gemini Spark. Whittemore’s case for choosing carefully is switching cost: “given how much context and setup and tool access and account access is going to be required… the cost of switching could be kind of high.” This article merges his four-question chooser with Claire Vo’s hands-on comparison of the products’ design primitives, and closes with the case against personal agents.
Key Takeaways
- Do not compare features. Whittemore, working from Every’s eight-agent comparison: “there is a very clear feature convergence.” They all connect to your apps, most can drive a browser, and all have controls.
- Ask four questions instead: work or personal; how much model control you want; UX or model quality; where your data goes and how you make it forget (table below).
- Only Hermes and OpenClaw let you bring your own model, run on hardware you pick, and edit memory directly. Every other agent runs in its maker’s cloud.
- Grok Bot’s model answer changed in October. Elon Musk posted that SpaceX “will use the best back-end model for any given task, including Claude Opus 5.5, Midjourney, Suno, and other leading APIs.” Until then it ran only on Grok models. Timing, model visibility and price impact are not announced (see Grok Bot).
- Speed and proactivity decide which agent Vo keeps. She calls Muse’s “killer feature” its speed. Duet threads took “10 plus minutes” and it “just wasn’t that proactive,” so it “fell to the bottom of my use list” despite being “approximately 100 times more polished than Dots was when it came out.”
- The primitives differ more than the capabilities. Muse has one main chat plus side chats, Duet has a main thread plus Slack-like channels, Dots has one “primary dot” for now, and Grok Bot has many single-job bots. Vo prefers Grok Bot’s shape: “I like to hire an agent for a job.” Brock Mesarich, after a few hours with Duet, agrees: separate agents are “a cleaner experience” than channels.
- Running several agents multiplies the setup. Whittemore: personal assistants “get more useful in proportion to what they know about you,” and most people will struggle to give even one of them full access. ^[inferred] Pick one primary and give it the access.
The four-question chooser (Whittemore)
| Question | What the sources say |
|---|---|
| 1. Work or personal? | Muse leans personal. Grok Bot “is being adopted by more people for work.” Dots looks work-aimed for now: it is paid-only and installs natively in Slack and Teams. Whether an agent uses iMessage/WhatsApp or Slack/Teams is today’s tell, but Whittemore “would be very surprised if in 6 months all of these agents didn’t just work in all of the messaging systems.” |
| 2. How much model control? | Three levels. (1) The maker’s own model: Dots uses GPT-6 Astra, Gemini Spark uses Gemini, Muse uses Muse Spark. (2) A builder-chosen mix: Grok Bot (the caption reads “a cursor managed mix”, leaning heavily on Grok models), and Instinct (its own trained core model plus unnamed third-party providers). (3) Bring your own: Hermes and OpenClaw only. Whittemore notes Muse is “the first AI product… where most of its users simply do not care what model is actually powering it.” |
| 3. UX or best model? | Hermes and OpenClaw lose some users to setup alone. Among the hosted agents the trade-off runs from Muse (“really dialed in consumer-friendly UX” on Muse Spark, “certainly not GPT-6 Astra”) to Dots (one of the strongest models, in a rougher product embedded in ChatGPT). |
| 4. Data and forgetting? | Hermes and OpenClaw run on hardware you choose; training depends on the model provider you plug in. Gemini Spark “currently requires you to allow it to train their models.” Dots, Grok Bot, Instinct, Muse and Poke let you opt out, and Dots excludes business data by default. Memory: edit directly in Hermes and OpenClaw; fix through chat in Dots, Muse, Gemini and Grok Bot; Instinct and Poke can only delete everything. |
Tie-breakers. Where your information already lives (ChatGPT, Google, Meta or Cursor), restrictions outside the US (“especially if you live in Europe”), and price: to experiment without a new bill, use what you already subscribe to or one of “the handful like Muse that are free to start.”
His own quiz result (illustration only): work use, no model preference, proactive “text or call when something needs me”, code in a terminal. Grok Bot came first at a 68% fit, Dots second at 62%. The quiz’s stated trade-offs: Grok Bot will not let him bring his own model, and “it’s not going to reach out to me. It’s all about scheduled or triggered tasks.”
Product notes
Launch order, per Brock Mesarich: Grok Bot in August, Muse in September, Dots “a couple weeks” before Duet, and Duet on 2026-10-21.
Muse (Meta)
- Status is the core design decision. Alex Cornell, Muse’s design director: “just simply having that status I think has been one of the most important decisions that we made.” It uses progressive disclosure in levels: “Jollybot is working” → “researching pumpkin patches” → a list of actions → “the curl and all the code”, which is “only there if you need it.”
- Acknowledge, then work. For a hard request it sends a reaction or a plan first. “Like a person… they’re not going to just ghost you.”
- Generative UI, but “on the rails.” Fixed primitives (avatar, status, ideas, feed) are filled with generated content. The Ideas surface exists to show capabilities users don’t know about, personalised from past chats.
- Field test. NYT reporter Eli Tan (relayed by Whittemore) had Muse fill in forms, pay a parking ticket from a photo while “asking for approval before making the payment”, handle dental insurance by phone, and bargain with Facebook Marketplace sellers. His skepticism moved from usefulness to privacy.
- Limits reported. Sacha Malik: good at immediate tasks, “terrible for long-term tasks” because it “gets the day wrong” on reminders. Sam Lessin: “Hilton site put my agent in a doom loop. Third retry I gave up.”
- Traction. Whittemore reports 3M weekly and 1M daily active users a little over three weeks after launch, against about three months for Codex to reach 3M WAU. It sat at number one on the App Store.
- Team and scale, early November (added 2026-11-09). Nate B Jones reports that 10 million people set up Muse in six weeks, and relays Alexander Wang (in an October interview with David Senra) that fewer than 200 people worked across the model and product and that the team tested and improved more than a hundred specific model behaviours before launch. Jones calls Muse “not the smartest agent on the market” but the most polished consumer AI app, says it is expensive to serve, and credits his own Muse with cancelling $1,285 a year of subscriptions (
raw/You_ve_Never_Lost_19_Billion._So_Why_Are_You_Playing_It_Safer_Than_Zuckerberg.md; see Muse and the Case for Another Swing). - Surfaces. Muse for Mac with computer use and a real-time avatar, Ray-Ban glasses support, and the Muse Charm device. Open-source Muse firmware for ESP32 plus a Linux SDK, so “any device with a Wi-Fi connection can now act as a Muse gadget.” iPad support, connectors for GitHub, Asana, QuickBooks and Canva, and Windows on the roadmap.
Dots (OpenAI)
- One primary dot for now. Sottiaux: “we’ll release the ability to create more than one”, each with a role (he wants one for monitoring Twitter). Vo: available on Pro, Business and Enterprise.
- No model picker. “There’s just no configuration. You just talk to it” (Sottiaux).
- Architecture (first-party). “The harness does not run on the machine.” A dot has its own computer, can be given your laptop, and “over time, you might connect it to like 10 different devices… a little bit like an octopus.” You can call your dot by phone and install it in Slack (Vo).
- Early verdicts. Vo: the work is good (“the bones of your dot… are good”) but the split between dot, chat and Codex threads is confusing. Analyst Max Weinbach’s expenses test (relayed by Whittemore): “Grok bot made a mess… Dots did it after the first time.” Critic Habibi Slap: “a plumber in a house where you’re not allowed to touch the pipes,” without “sufficient access” to ChatGPT and Codex data.
- A Hermes user’s first 24 hours (Sharbel, 2026-09-30). Source:
raw/OpenAI_Just_Launched_Dots._Here_s_The_Truth_About_It..md(https://www.youtube.com/watch?v=HCttXibjt-c, 12:20). Not available in his EU region at launch; he got access another way.- Setup: open the ChatGPT desktop app, name the dot and pick an avatar, connect apps (OpenAI says 4,000+). It produced a 10-page deck comparing itself with Hermes and Grok Bot.
- What he credits: it reads documents you connect and proposes tasks unprompted (“extremely creepy” as well as useful); background work is read-only by default, so it drafts but cannot send, post or spend without approval; team mode shares a dot’s artifacts and folders with colleagues’ dots, which he says Hermes makes hard (he zips folders between teammates).
- What he faults: region limits; chat is nearly free but tasks come out of your usage, on the same day the $200 Pro plan’s allowance halved and Pro 500 arrived; day-one bugs and over-messaging. The phone call is “nothing different from a GPT voice mode session.”
- His verdict: little that Grok Bot or Hermes does not already do; “the magic is that OpenAI just put this in front of over a billion people at once.” Worth trying if you already pay for Pro; builders who want to own their setup lose nothing by staying on Hermes or OpenClaw.
Grok Bot (SpaceX AI)
- Many named single-job bots, each with its own cloud computer. Full playbook in Grok Bot.
- Going multi-model (October). Musk’s post, read word for word: “Going forward, SpaceX will use the best back-end model for any given task, including Claude Opus 5.5, Midjourney, Suno, and other leading APIs.” Whittemore adds Musk’s follow-up that most requests will go to “a lightning fast version of Grok 4.8 when that comes out.” Some users asked for a Grok-only opt-out.
- Runs in Teslas through the existing Grok voice interface (Whittemore).
- A two-month daily user’s setup. Brock Mesarich has it reply to “basically all of my emails”. A main agent works with his invoice and YouTube agents and “has all the context from all of them”, so his conversations “don’t bleed together”.
Claude Duet (Anthropic)
- Phone-first research preview aimed at “business user, SMB, entrepreneur, busy PM.” It inherits your Claude plugins and artifacts and has task lists and channels. Full product details in Claude Duet.
- Vo’s complaints: slow, not proactive enough (“there are only so many three-point morning briefings that can pull me back in”), no desktop app, no local skills, and single-player only.
- Brock Mesarich’s live test (2026-10-22, phone only): what he rates above Grok Bot is the Make section, which holds every artifact on the account including Claude Code output, and the Tasks view, which adds work he never asked for, such as an important email queued for tomorrow. Duet built a dashboard from his Supabase data, and another from Drip once he had connected Drip through Zapier MCP, each in about a minute. He stayed on Grok Bot. Details in Claude Duet.
Hermes, OpenClaw, Instinct, Poke, Gemini Spark
- Hermes and OpenClaw are the bring-your-own-model, self-hosted end of the field. See Hermes Agent and Two Shapes of the Personal Agent.
- Instinct mixes its own model with third parties and can only delete data wholesale. See What Personal Agents Keep.
- Poke and Gemini Spark appear only in Whittemore’s matrix in these sources.
Update 2026-11-01: a benchmark and two operators’ splits
- Claire Vo’s nine-task benchmark (
raw/I_tested_4_AI_agents_on_9_real_tasks._Here_s_who_won..md, https://www.youtube.com/watch?v=hnCoQ5wB-GU): Grok Bot 93%, then Muse, then Dot, then Duet at 79%. Grok Bot won thumbnails (Figma), inbox to-dos and shopping; Muse won the short video, family newspaper and homework explanation and was fastest on 4 of 9 tasks; Dot and Duet were best at the coding task; Duet won the launch video. Her split: “if you’re picking work jobs, probably give it to Grockbot. And if you’re picking life jobs, probably get it to Muse.” Method and per-task table: Claire Vo’s Personal Agent Benchmark. - Nate B Jones’s split (
raw/The_Future_Of_The_Internet_Isn_t_Open_For_Your_Agent._Muse_Dots_And_Claude_Duet_Will_Find_Out..md, https://www.youtube.com/watch?v=c3HHFL_lsLM): Grok Bot for harder builds where bots must share discoveries (his chief of staff lives there); Dots for work that needs his Codex context, coordinating calendars and email (“starting over is something that I don’t have to do anymore”); Instinct for travel and texting on the go; Muse for preliminary shopping, gifts and bookings, including paying through Link; Duet for proactive reminders. He rates Dots’ and Duet’s proactive alerts useful “50 to 75% of the time”; Dots caught a double-booked flight he then cancelled for a refund. He adds that “normal people pick one,” and expects Muse to lead on consumers while Grok Bot, Duet and Dots lean “prosumer and work.” - Nathaniel Whittemore’s split (
raw/How_I_m_Actually_Using_Claude_Duet.md, https://www.youtube.com/watch?v=HyEfRBW10Kk): he uses his Dot as a “master holder of context” but finds it “very weak on proactive work” and flat for strategy; Duet is the proactive one (details in Claude Duet Early Reviews). He keeps Grok Bot “just enough engaged” for its X integration and skips Muse as too personal. His view on choosing: features that work in one agent will be in all of them “within weeks, if not days.”
The case against (and the variable to watch)
- Ben Thompson (relayed by Whittemore): “consumers do not care about being productive… people love shopping.” His assistant of ten years “does not come within an inch of booking my flights for me.”
- Counters quoted in the same episode: Simon Taylor: “People are lazy. They want to delegate schlep work”, such as booking a child’s cycle class or a dinner time that suits everyone (“That’s not shopping, it’s buying things”). Joe Gunter: anyone who doubts demand for outsourced admin forgets “there are 2 billion mothers in the world.”
- Habibi Slap on flight check-in: verifying the agent’s work costs about as much as doing it yourself.
- Proactive versus reactive. Riley Brown described a poker table of non-tech professionals who said of a personal agent app (the caption name is garbled), “you don’t have to ask it to do stuff… you just connect it, and it starts working for you.” Vo reached the same conclusion from the other direction: Duet lost her because it never pushed anything.
- Ethan Mollick, “The Dot and the Swarm” (read by Whittemore): “they are finding my mistakes rather than having me identify theirs,” for example an error on a town permit and an airline credit about to expire. “The more important thing is what you no longer have to tell them.”
Try It
- Answer the four questions in writing before you install anything. If two answers pull opposite ways (for example “personal” and “best model”), decide which matters more; Whittemore’s quiz flags that tension rather than averaging it away.
- Run one eval task across your shortlist. Vo uses shopping (shoes, game tickets); Weinbach uses a week of expenses from email and Drive. Time each agent to a closed loop, not just to a correct answer.
- Check the memory and training settings on day one. Opt out of training where offered, and find out whether you can edit memory or only delete it.
- If you build an agent, copy Muse’s status ladder. Show “working” by default, what it is working on one tap down, and the raw tool calls at the bottom.
- Acknowledge long requests immediately with a reaction or a one-line plan, then work.
Open Questions
- Grok Bot multi-model details. Not announced: when it ships, whether users can see or pick the model, whether Opus-routed work costs more against usage limits, and whether data sent to a third-party model can be blocked.
- Whittemore’s matrix is a snapshot. It comes from Every’s comparison and his own reading in mid-October. Training and memory settings change; check each vendor’s current privacy page.
- The Muse traction numbers are Whittemore’s report, not a Meta filing.
- “Cursor managed mix” is the auto-caption for how Grok Bot chooses models. The exact wording is unverified.
- Dots team dots and multiple dots were “not released yet” at DevDay; no date given.
Related
- Grok Bot — Running a Business on an Agent Team — the single-job-bot playbook
- Claire Vo’s Personal Agent Benchmark — four agents on nine tasks
- Claude Duet — Anthropic’s product details
- Two Shapes of the Personal Agent — self-hosted versus hosted
- What Personal Agents Keep — disconnect is not delete
- Hermes Agent — the bring-your-own-model option
- OpenAI DevDay 2026 — the Dots launch in context
- Microsoft Autopilot — OpenClaw-style agents arriving at work
- The Forward-Deployed Engineer Playbook — why enterprises treat these as “sidekick” agents
- ChatGPT Work for Marketing (Grace Leung) — field report on Dot running a marketing agent team