Ongoing topic · Compiled October 5, 2026

Personal AI partners: ongoing work and human decisions

Connect dots, Muse, Grok and Instinct with Duet’s proactive briefs, 1Password login approval and Lingguang 2.0’s household tasks. Follow context, accounts and actions, separating the decisions that return to a person.

A detail of the DevDay stage: Sam Altman stands beside the dots name and four colorful characters.
dots appears at DevDay 2026 as an ongoing assistant. This view enlarges the stage; open the image for Simon Willison’s full photograph.Simon Willison · DevDay photograph · Original source ↗View full image ↗
Read by questionContents20

01A change worth following together中文

From answering a question to carrying unfinished work forward

Can an interview transcript become clips, show notes and social copy, with later edits carried across the set? Can school notices turn into a family calendar and a supplies list? Can recurring research and email preparation continue after one handoff? dots, Muse, Grok Bot and Instinct belong in one view because each explores assistance that keeps context, enters existing tools and carries work forward.

OpenAI dots connects Astra to a cloud computer and ongoing tasks. Meta Muse produces plans and checklists while extending toward glasses and Charm. Grok Bot gives recurring work a role and cloud workspace. Instinct uses messages and calls for everyday requests. Their interfaces make delegation feel familiar, while the work they emphasize and the outputs they leave differ.

This topic starts with official examples and outputs, then connects firsthand accounts and later projects, distinguishing software characters from device interfaces. The table maps uses and entry points; task records explain how well they work.

October 5 correction: dots here means the ongoing assistant OpenAI introduced at DevDay. The initial version used material about an unrelated iPhone app with a similar name; it has been replaced.

Four personal AI interfaces: the work they take on, where it runs and where results appear. The overview was checked October 5; Instinct group chats were added October 6.
Dimensiondots · OpenAIMuse · MetaGrok BotInstinct
Example workTurn a transcript into creator materials and carry revisions forwardTravel plans, school information, shopping and checklistsResearch, outreach drafts, team recaps and recurring workAirport rides, home repairs and dropped threads
How to reach itChatGPT messages or calls; Slack and Teams can be connectedApp, web and WhatsApp; glasses and Charm are later interfacesMobile or desktop messages to a Bot with a name and roleFamiliar messages and calls
Where work happensAstra and a cloud computer and browser; optional local connectionMuse Spark with a personal Secure VM and browserA persistent cloud computer, files, browser and connectorsConnected apps and devices provide context for everyday tasks
Current accessEligible Pro, Business Premium and admin-enabled enterprise accounts; create on desktop firstOfficial U.S. rollout; glasses and Charm details are still to followA paid Cursor plan or a linked individual SuperGrok subscriptionThe site offers messaging; group chats start with early-access users. Regional eligibility is not stated
Sources and notes

02dots · Carry a job forward中文

From a transcript to a set of materials, then revisions across the set

OpenAI’s dots illustration shows Iggy selecting transcript moments, preparing show notes and social posts, then revising materials after feedback.
Official creator scenario: one transcript becomes several materials, with edits carried into related drafts.OpenAI · dots product illustration · Original source ↗View full image ↗

In OpenAI’s creator scenario, Iggy finds moments in an interview transcript, prepares show notes and social posts, then carries feedback across those materials. The official illustration turns one copywriting request into continuing creative work: tone, audience and accepted edits inform the next batch.

Astra, a cloud computer and browser, and relevant conversations, memory and preferences support this continuity. dots can research, work with files and run software. Create it on desktop, then continue in a supported mobile app. Local projects require the connected computer online with ChatGPT open. OpenAI also relays an early tester’s missed invoice being prepared and sent after approval; that is a vendor-reported case.

SLIDEFACTORY’s Mark Nguyen reports roughly five hours of early-October agency work: dots found calendar conflicts, joined Slack, summarized a PDF and, through the agency’s own tool, completed crawls of 44 and 111 pages on two sites. Email and rescheduling remained unfinished, and the first scheduled report had not arrived. The completed outputs and open tasks show what the first day established.

Creer also reports using existing Markdown, JSON and handoff files to resume two local projects and continue edits and tests separately. The DevDay topic follows how saved material becomes a starting point for the next step.

Sources and notes

03Muse · Useful outputs and portable interfaces中文

Travel, school and packing lists connect to one portable assistant

The checkable packing-list card enlarged from Muse’s three official artifact examples; the full image also includes an India itinerary and school information.
The packing checklist enlarged from three official artifact examples: a checkable result from a chat task. Open the full image for the travel and school pages.Meta · Muse product team · Original source ↗View full image ↗
Zuckerberg holds Muse Charm at Connect, with its avatar enlarged on the stage screen.
At Meta Connect, Muse gains a pocket-sized voice interface. The stage device is a physical prototype; launch details are to follow.Meta keynote image via Shacknews · Original source ↗View full image ↗

Muse’s team published an India itinerary, a school information page and a checkable packing list: material from a conversation becomes something to reopen and use. The team also describes extracting dates from school emails and preparing supplies. These official outputs and accounts connect the idea of a personal assistant to recognizable everyday work.

Charm and glasses change how a person hands off a task. At Connect, Mark Zuckerberg held a physical Muse Charm prototype, with its avatar on the screen. Meta also announced Muse for AI glasses: looking at a shelf product, a notice or a school-supplies list could provide context without describing it. The announced Charm prototype is designed around pocket-sized real-time voice interaction; launch details are still to follow.

Software, devices and makers connect here: Muse works in a personal Secure VM and browser; hardware makes the same agent easier to reach. The ESP32 firmware and Linux SDK in Muse Gadgets offer makers routes through buttons, displays and sensors. The next useful evidence is a working project showing why its particular form improves an interaction.

Sources and notes

04Ongoing agents · When a person is needed中文

Claude Duet: what keeps moving, and which decisions come back to you?

Claude’s official release-video cover: an orange Duet icon and Claude Duet name on a pale lavender background.
Official release-video cover identifying the product, not a task-test screenshot. Open the source for the full announcement video.Anthropic / Claude · Official release-video cover · Original source ↗View full image ↗

An Anthropic-selected tester says Duet prepared a client brief, noticed a clash with her flight and proposed rescheduling. The useful idea is noticing a conflict from existing context, then returning the decision to the person. This is a vendor-selected account, not our test or a guarantee of timely detection.

Launched October 21, Duet uses existing Opus 5.5, Claude memory, skills and connected apps. A continuing text/voice thread delegates work to its cloud computer. Like dots and Muse, it can continue away from a device; its particular starting point is an existing Claude account and professional work. This is not a new model.

Official rules separate preparation from action: sending, paying, signing or deleting requires an explicit approval; apps can be disconnected. The described Link purchase flow also separates item confirmation from spending approval. Connection, preparation, authorization and completion remain separate checkpoints.

Access is a US Pro/Max waitlist preview, excluding Team and Enterprise. Standalone iOS/Android apps and web access do not establish admission for every subscriber. Background checks share the plan’s usage limits and run every few hours; users can specify what merits a notification and how often.

Personal errands are weaker in this preview. It may misread urgency or miss arrivals between checks, and cannot yet show its browser. For a meaningful comparison, trace one task from the original message through a proposed action, readable approval and final record. Those are our suggested observation points, not a completed cross-product test.

Sources and notes

05Account access · Sign-in and actions中文

1Password: why approving a sign-in still leaves a purchase decision

Official 1Password illustration: allow Muse to sign in, with Authorize and Deny choices.
Complete official Muse approval illustration, not our authorization, payment or security-test record. Eligibility and rollout are stated in the text and sources.1Password · Official product illustration · Original source ↗View full image ↗

A party-planning agent may need an invitation service and a store. In its October 22 example, 1Password connects that workflow through an approved saved login, filling the site without giving the agent the password or one-time code. These are publisher examples, not our completed orders or account tests.

Separate connecting an agent, approving a login, and approving what happens inside the account. The platform FAQ says a signed-in agent may read information or act, depending on the site and agent. Muse’s page separately describes confirmation before purchases or emails. Keeping a password out of model context does not by itself establish correct actions.

Rollout is integration-specific: Muse and Grok Bot are rolling out over the coming days; Duet, ChatGPT Work, dots and Instinct remain coming soon. Partner logos do not establish availability in your client. July’s Claude browser integration does not establish access to the new Duet connection.

The new personal-agent connections are included in Individual and Families plans on 1Password.com, not .eu or .ca accounts. Grok’s page also excludes Business and Team Starter Pack. No additional integration fee does not make the agent free or remove its own eligibility conditions. Disconnecting credential access should not be assumed to log out every existing website session; check the agent and site separately.

Sources and notes

06Origins and landings · Two new pieces of evidence中文

dots’ capabilities came from coding tools; Lingguang 2.0 connects bills and class groups

In a Disrupt interview published October 26, Alexander Embiricos, who leads OpenAI’s enterprise product team, explained where dots came from: it repackages capabilities from Codex, the coding app OpenAI released at the start of the year, and from ChatGPT Work, which itself grew out of agentic coding tools. He said much of dots’ work was already possible in Codex, just with more tending — his wife once left her computer running overnight so her agents could keep working. The interview clarifies the “cloud computer plus ongoing tasks” pattern this topic keeps returning to: infrastructure built for developers first, later given a conversational shell.

The interview also grounds “proactive” in mechanism and limits. dots runs “proactive research”: on a cadence it reviews what it knows about you and your conversations, researches how to be more helpful, and proposes actions. Both cited examples stop at noticing and proposing: Embiricos’ dot flagged a calendar conflict only three minutes before the event, and he expects that timing to improve; an engineer at a drone-delivery company said his dot detected and analyzed an outage, then proposed fixes, before the engineer noticed. Pricing and privacy are concrete too: dots requires ChatGPT Pro or higher, starting at $100 a month; privacy controls follow ChatGPT but are not granular — the app refuses read-only access to cloud drives — and OpenAI is in conversations with payment processors about letting agents spend money. Embiricos himself keeps a human assistant, because corporate partners want a person.

The same infrastructure reached Chinese everyday services on October 27. According to an IT之家 report, Ant Group released Lingguang 2.0 that day, describing a shift from “responding to instructions and finishing single tasks” to “proactively finding needs and continuously managing them”: an independent cloud host supplies cross-session memory and long-horizon scheduling, and tasks keep running after the user leaves the conversation, following new information as it arrives. The scenarios listed sit on the household and consumption side: sorting Alipay bills, surfacing forgotten auto-renewals, alerting on flight and hotel price drops, turning teachers’ notices in DingTalk class groups into to-dos and formatted documents, plus long-running matters like pet care, childhood vaccinations, family anniversaries and vehicle inspections. Services span Alipay, DingTalk, Feishu, email and Mijia, with MCP and CLI tools available.

Security and availability need separate checks. The report says Lingguang uses Ant’s APASS as its account and security foundation, with authorization keys held apart from the agent — the agent never touches the keys directly, and users can view, adjust or revoke access at any time. That is the same question as the previous chapter, where 1Password separated approving a sign-in from approving sensitive actions. Availability deserves care: when Lingguang 2.0 entered testing on October 12, Ant told reporters it was “still in a small-scale internal test,” and the new report gives no public entry point, eligibility or pricing. Until a verifiable first-party entrance appears, this is an announced direction rather than a product ordinary users can try; this site has not logged in or tested it.

Placed back on this topic’s main line, the two reports add two things. First, dots’ origin shows these personal assistants are not a new species appearing from nowhere; they are developers’ background-execution capabilities — cloud computers, long tasks, tool connections — wearing a conversational shell. When choosing one, it helps to ask what its “previous job” was. Second, Lingguang 2.0 shows the same infrastructure can plug into very different life services — Alipay bills, class groups, Mijia — not just calendars and inboxes. Two threads to follow: whether the needs an agent “discovers” are needs you recognize, since a three-minute heads-up says detection timing is still hard; and where results land once you leave the conversation and who reviews them — the same question raised by dots’ project accounts and Grok Bot’s outreach log.

Sources and notes

07Grok Bot · An ongoing job中文

Give it a role that connects research, drafts and recaps

SpaceXAI’s official marketing-demo task card: draft launch copy in Figma and Notion and recap the campaign for the team.
The official demo task card connects drafting and a team recap; this frame does not show a finished publication.SpaceXAI · Original Team Bots demo frame · Original source ↗View full image ↗

Grok Bot’s official marketing demo combines launch copy in Figma and Notion with a team campaign recap. Another official example starts from a weekly CRM list, adds web and Slack research, and leaves outreach drafts for review. The role lasts beyond one question, with material and outputs in existing work tools.

Give a Bot a name and responsibility, then message it from phone or desktop. It works with a persistent cloud computer, files, browser and connectors. Official documentation says cloud work can continue after the laptop closes, and Bots can divide work. A weekly job can therefore retain material, methods and corrections across runs.

James Swierczewski’s outreach account gives a concrete process: the Bot proposed a list, contact research, email drafts and a tracker, with final sending left to him. When tracker files were invisible, it moved records and drafts into chat. The account makes output visibility tangible. Grok Companions emphasizes character and voice conversation; this chapter follows the work-focused Grok Bot.

Sources and notes

08Added October 9 · Ongoing context for a person and team中文

Gemini's work agent: can context survive a new interface or model?

Google's work-agent illustration: one prompt box, Auto routing and tasks for a weekly report and product launch material.
Google Cloud's official interface illustration. Completed tasks are demo content; the FAQ's early-access conditions govern background delegation and interfaces.Google Cloud · original release image · Original source ↗View full image ↗

Google's October 8 work-agent release puts a weekly report, launch material and pricing analysis in one illustrated task list. It continues this topic's question: as work moves from a document to email, data and code, can context and revisions travel with it? Done is the official illustration's status, not our task result.

Google describes cloud execution with memory of tasks, business facts, methods and past actions, reached through existing tools such as Workspace, Microsoft 365 and Slack. Alongside dots and Grok Bot, it separates an ongoing work environment from the chat window. Outputs still need a place people can open, inspect and revise.

Gemini here names an agent product, with the underlying model a separate choice. The keynote describes Gemini and Claude orchestration, while the product FAQ checked October 9 lists Google models today and third-party choice in early access. This does not give every trial account Claude access or establish a new model for the general leaderboard.

The product page offers sales contact and a 30-day Enterprise Plus trial, while marking background delegation, mobile/desktop access and third-party models as early access. This differs from the consumer Gemini plans and October 9 schedule previously covered here; access and usage conditions are separate.

Compare this morning's Claude Google-file integration: it asks how to keep editing a file, while this product asks how context crosses tasks and tools. Our suggested trial is existing material, checking where outputs land and whether the next revision continues before expanding use. This reading uses official material without replicating business-account tasks.

Sources and notes

09From a spoken commitment to a next step中文

Bee turns conversations into to-dos. What does the next step cost?

Bee’s official wearing scene: a woman talking at a table wears a black Bee Frontier, with its body and button visible on her wrist.
Frontier keeps a conversation interface on the wrist. This wearing scene comes from its official product page. Frontier is available on Amazon.com; the magnetic phone attachment Eclipse is still coming soon. The image illustrates use without proving accurate capture or a completed task.Bee / Amazon · Official wearing scene · Original source ↗View full image ↗
Bee’s official suggested-to-do interface: confirm an engineer’s Tuesday availability, with Draft email and Create event buttons; a second item asks Sam to review drawings and specifications before a client meeting.
The official illustration carries a conversation into a next step: confirm availability, then use an email or event action. It shows task organization without establishing sending or a completed appointment. Suggested to-dos and integrations require Pro.Bee / Amazon · Official product illustration · Original source ↗View full image ↗

Forgetting what you just agreed to is the everyday problem Bee addresses. Its official interface turns confirming an engineer’s Tuesday availability into a suggested task, with buttons to draft an email or create an event. Conversation becomes material to revisit and continue in connected tools. Amazon announced two devices on October 15: the $74.99 Frontier is available on Amazon.com for wrist or clothing use; the magnetic phone attachment Eclipse starts at $99.99 and is still coming soon. The hardware changes how conversations are captured; the task illustrates what follows.

As checked October 16, the free tier includes conversation processing, summaries and 14 days of history. Suggested to-dos, app integrations, Alexa+ and Amazon.com connections, desktop access and unlimited history require Bee Pro: $11.99 monthly or $99.99 annually on its website, with 30 days included for new devices. Alexa+ can retrieve summaries and to-dos; Amazon.com can suggest mentioned products. A remembered commitment, a prepared action and a completed task remain separate outcomes: a draft, event button or recommendation does not establish sending, attendance or purchase.

Amazon says audio is transcribed in real time and discarded after processing, leaving encrypted text, summaries and to-dos under device-controlled keys. Processing also uses Amazon’s isolated cloud environment, so this is not a wholly offline system. A button and indicator control capture; call capture prompts users to obtain participants’ permission, and service connections involve chosen sharing. Anyone needing original audio for verification should know it is not retained. This continues the topic’s question of where usable results remain: inspectable memory, chosen connections and actual completion determine longer-term value. BitShovel reviewed official material without wearing or testing the new devices.

Sources and notes

10Team knowledge · Context for the next task中文

Stop piecing the project together for every question

Microsoft pipeline: SharePoint, Outlook, Teams, Word, OneDrive and PowerPoint content passes through Work IQ into contents, entities, topics and summaries, forming a Markdown knowledge base for people and agents.
Microsoft's overall pipeline diagram, obtained intact from Vesa Nopanen's credited republication. Sources on the left become organized relationships and shared knowledge. The design does not mean all curation and integration features are available, nor document a BitShovel enterprise-account test.Microsoft · credited republication by Vesa Nopanen · Original source ↗View full image ↗

Following Gemini's cross-interface context, Microsoft's October 14 SharePoint preview shows how shared context can be formed: people contribute emails, messages, meetings or site files, and Copilot organizes them through Work IQ into linked Markdown knowledge pages for people and agents. This is project context, not automatic sharing of private chats.

Microsoft reports up to fourfold response speed and 30% better token efficiency in early internal tests conducted in October 2026, not a universal Copilot gain or a guarantee of 30% fewer tokens per task. Our reading: organizing relationships in the material can change the work needed on the next question.

Access requires the Frontier preview and usage-based billing, with admin controls; Autopilot has a separate rollout. Automated ongoing curation, more Cowork/Copilot Studio access and CLI integration remain planned. Our suggested first check is how pages preserve support and correct old decisions. We have not reproduced this in an enterprise tenant.

Sources and notes

11Continue the same work away from the desk中文

How does an AI task on your computer continue while you are away?

Cursor's official desktop Remote Control settings, including keep-awake and a paired iPhone.
Original image from Cursor's announcement. It shows desktop remote-control settings; Connected describes the official example, not a BitShovel device.Cursor · official changelog image · Original source ↗View full image ↗

Cursor's October 6 Remote Control release lets its iOS app show agents on your computer and accept follow-up messages. This adds another starting point to mobile delegation: work has already begun in your project, and the same conversation continues away from the desk.

The detailed help page checked October 7 separates the agent loop in Cursor's cloud from terminal commands, file edits and tests on your computer. That computer must stay awake and online. Context and tool results reach Cursor, so local execution here does not mean fully offline inference.

Grok Bot's persistent hosted computer offers a useful comparison. Remote Control continues with your existing local files and test setup; Grok Bot starts from a hosted workspace and can separately receive permission for local actions. Both have mobile interfaces, while the execution location determines how work continues.

A mobile interface does not determine the execution environment; based on official documents checked October 7.
QuestionCursor Remote ControlGrok Bot
Starting pointAn agent already working in a project on your computerA Bot with a role and persistent workspace
Files and toolsTerminal, edits and tests on your machine; agent loop in the cloudA hosted workspace; local actions have separate permissions
Before leavingAn awake, online computer and the account/team access settingsResults remain visible in the chat or workspace for review
Sources and notes

The announcement and detailed help differ on default enablement and Cloud Agents requirements. This section retains that distinction and uses the help page's more specific runtime description. Check current documentation and account settings for access. BitShovel did not pair devices or run a task.

12Xiaoyi · Retrieve material from your computer中文

Find a file on your computer from the phone

Huawei’s three Chinese-language Xiaoyi screens show a request for health-check reports on the computer, two documents from 2026 and 2025 with a question about which to send, then two file cards and a transfer-confirmation reply after the person chooses both.
Huawei’s official Chinese interface example, preserved in full, not a BitShovel device test: request reports from the computer, receive two choices, select both, then see file cards and the transfer-confirmation message. File-list and card sizes disagree; they are not measurements of size, speed or transfer accuracy.Huawei · Original figure from the Xiaoyi cross-device file guide · Original source ↗View full image ↗

Huawei’s guide shows a phone request for reports on a computer. Xiaoyi lists two documents, asks which to send, then displays file cards after the person chooses both. This is an official interface example, without a BitShovel device test or a guarantee for arbitrary files.

This route is for HarmonyOS phones and computers meeting the original guide’s requirements: HarmonyOS 7.0.0 or later, software 7.0.0.105 SP10 or later, the same Huawei account, screen-lock passwords and network access. Authorize the phone in computer Xiaoyi and open its expanded window first. Check device and version eligibility in the guide. It also lists translation, summaries, mind maps and packaging; review the resulting files for quality.

A powered-off computer cannot respond; sleep requires power, and a reboot requires one desktop login. After 30 days without device connections, reopen Xiaoyi’s expanded window. Compare this with Cursor continuing a local project and Muse’s hosted workspace: the phone is an entry point; files, authorization and the running computer still determine the handoff.

Sources and notes

Based on Huawei’s guide read October 12, 2026. The support page does not expose an original publication date. Its file-list and card sizes disagree, so the image illustrates the sequence rather than transfer accuracy or speed.

13Manus × OPPO · Phone material and finished results中文

How does material on a phone become an itinerary or a document?

The English Manus/OPPO photo illustration previews three albums with 10, 32 and 16 selected photos, Cancel and Confirm buttons, and an AI-generated demonstration label.
Complete frame at about 6.5 seconds of the original 11.3-second GIF: preview, adjust, then confirm. The English illustration is explained in the text; its numbers and timer are not BitShovel measurements.Manus / OPPO · Original frame from the official photo illustration · Original source ↗View full image ↗

On October 13, Manus announced Agent Mode integration with OPPO AI Mind Pilot, planned for Find X10 with ColorOS 17. Its travel example gathers bookings across phone apps and adds online information to an itinerary; files and photos can also inform documents. Our interest is how system access can reduce the material-gathering step before an agent organizes a result.

The photo illustration previews portraits, food and views for a person's confirmation before showing completed albums. Its 10, 32 and 16 photos are demonstration counts, not accuracy or speed measurements. The screen labels itself AI-generated for demonstration. BitShovel reviewed the official material without reproducing the task on a phone.

Manus describes local understanding and file organization through on-device models, with permission requested before external processing. Our comparison with Xiaoyi's computer-file retrieval concerns where material lives and how it becomes an output. These separate systems are not a controlled comparison or proof of equivalent access.

OPPO's global page specifies device, region, version and service limits, limited free usage, then a ChatGPT or Manus login to continue. A ColorOS upgrade does not guarantee every feature at launch. Check actual availability, then whether the right material was found and the output can be revised.

Sources and notes

14Doubao phone assistant · Cross-app actions中文

A phone assistant can tap apps. Who lets it act?

Editorial SAEP map: user confirmation, system identity and risk, and app rules jointly determine whether an action continues, returns to the user or stops.
An editorial map of the official rules. Three checks contribute to a decision; permission to continue still requires app checks and does not mean the task is complete. This is not a phone interface or a BitShovel test.BitShovel editorial map · official Doubao SAEP guide · Original source ↗View full image ↗

On October 14, Doubao announced a gradual expansion of third-party screen automation after the SAEP notice period. Taobao, WeChat and Meituan remain in the official refusal list opened that day. It covers Nubia M153 and NaviX Ultra and is updated on Wednesdays. Explicit refusals remain excluded, and developers may change their declarations later; this is not universal app control.

The refusal concerns an agent’s screen-based taps and input, separately from installing or manually using the app. SAEP allows app-, page- and action-level rules, such as restricting login screenshots, input or final publication. Access to one app does not imply permission for every step.

User authorization, system checks and app policy jointly determine an action. Authentication, payments, deletion or publication may require human takeover or refusal. Permission to continue still leaves app-level checks; a confirmation covers the displayed step. ByteDance’s open demo declares policies and queries identities and its own logs. It is not an assistant for controlling arbitrary Android apps.

Alongside Manus × OPPO, this shows different access questions: local material and external sharing versus third-party screen automation. These separate systems are not a controlled comparison. For a task, check how material is obtained, which apps and steps can run, where you take over and what result remains. BitShovel read the announcement, expanded list and integration material without reproducing a phone task.

Sources and notes
  • Doubao: SAEP announcement and refusal list ↗Official developer page · October 14Opened the developer page through the homepage SAEP navigation and expanded its list. Gradual access, changeable refusals and M153/NaviX Ultra scope; differing media counts are not a fixed supported range.
  • Doubao: official SAEP integration guide ↗Official technical documentationAuthorization, system rules and app/page/action policy combine. Interface and enforcement support depends on the target system; documentation is not a phone test.
  • ByteDance: original SAEP-demo repository ↗Official open integration demoIts own policy, identity and log queries. APK installation compatibility is separate from runtime SAEP support.

15Instinct · Ask for help in a familiar way中文

Hand off a ride or a home repair by message or phone

Instinct’s official examples are familiar: arranging an airport ride, booking a handyman or following up on a dropped thread. Its interfaces are messages and calls, with an emphasis on understanding current work and personal priorities. These product descriptions make the intended everyday situations clear.

It connects relevant context from email, messages, screen, audio, location and devices, then uses a phone and computer to pursue the task. It may initiate a message or call. Familiar communication becomes a place for delegation and follow-up. Later experience can show whether an agreement survives and whether the assistant returns at the useful moment.

The public page does not state regional eligibility, and available work depends on connected services. For bookings or contacting others, establish who gives the final confirmation. Instinct’s terms also require a separate request to delete indexed information after disconnecting a service. These conditions affect ongoing use.

Sources and notes

16Instinct · From personal requests to shared plans中文

An AI companion joins friends making a plan

Instinct's official group-chat demo turns a Saturday hike discussion into two route options.
A frame at about 23 seconds: the group sees one set of options in chat. Messages and people are part of the product demonstration.Noah Shinn / Instinct · Official demonstration · Original source ↗View full image ↗

Instinct's group-chat announcement appeared early October 6, Beijing time. Early-access users can add a new Instinct serving the whole group. The demonstration turns a weekend suggestion into two hiking options, with the conversation continuing around the shared plan.

The interesting work is coordination: one person's preference, another's availability and a plan everyone can inspect. A personal assistant starts from one person's context; a group brings several needs into one conversation. This is our reading of the demonstrated scenario.

Follow what happens when times change, people join or the group moves toward a booking. Grok Team Bots provides a related reading about shared methods. Once AI enters an existing conversation, its outputs need to stay visible and usable to the participants.

Two handoff scenarios and the work that needs to continue.
What to inspectPersonal handoffShared plan
Where conditions come fromOne person's context and requirementsParticipants' times and preferences
Where the result staysSomewhere the person can inspect and reviseA plan the group can keep discussing
What to follow nextHow new material returns to the original jobHow changes return to the shared arrangement
Sources and notes

17Muse · Where outreach stands中文

Hand off the calls you avoid — will the other end take them?

Editorial diagram. Top, a timeline of the widening channel: from September a small set of early adopters (unlocked only by asking the AI); reported Nov 10, Apple Watch and Wear OS entries (voice, texts, dictation); announced Nov 12, calling opens to all paid subscribers in the US; next, Muse email addresses for free users. Middle, “the two ends of the same call”: dialing out — insurance, cable support, the DMV, restaurants without online booking, with Muse past 10 million users and plans at $20/month (500M tokens a week) and $100/month (3B tokens); picking up — many businesses and receptionists hang up once they identify an AI (as TechCrunch reports), voices mimic humans better so telling them apart gets harder, and Instinct and Google Gemini place such calls too. Bottom, boundaries as reported: paid US subscribers only, Meta’s subscription pages not publicly readable, and this site registered nothing, placed no call, verified no outcome.
An editorial diagram from TechCrunch’s Nov 10 / Nov 12 reports: the widening channel, the two ends of a call and boundaries as reported. It is our editorial explanation, not a product interface or call record; pricing and scope follow the text and sources.BitShovel editorial diagram · based on TechCrunch reporting · Original source ↗View full image ↗

On November 12, Alexandr Wang — Meta’s Chief AI Officer and head of Superintelligence Labs — announced that Muse can now place calls to U.S. businesses on your behalf, open to all paid subscribers from that day. Since September the feature had been limited to a small set of early adopters, unlocked only if you asked the AI for it. The official examples are the calls people would rather not make: insurance, the cable company, the DMV, or a restaurant without online reservations. Pricing and eligibility follow TechCrunch’s report of the post: subscriptions start at $20 a month (500 million Muse tokens a week), with $100 a month (3 billion tokens) for heavier users.

Calling is not an isolated update but another widening of the same delegation channel. On November 10, TechCrunch reported Muse on Apple Watch and Wear OS, taking tasks by voice, text or dictation — about a month after the iPad app arrived; the same report notes Muse passed 10 million U.S. downloads within 40 days of launch, the 12th app in App Store history to reach that milestone. Wang also said the recently added Muse email address will soon roll out to free users. Entries now span web, iPad and watch; the pace of that expansion matters more than any single feature.

But dialing out is half the call; the other end is the other half. TechCrunch reports that many businesses and receptionists hang up as soon as they identify an AI caller, and as voices mimic humans better, telling a customer from an agent will get harder. Instinct and Google’s Gemini place such calls too — multiple capability sides are building on this channel at once. This site registered nothing, placed no call and verified no outcome; the hang-ups are a media-observed pattern, not our observation. It returns to this topic’s standing question: an agent must first be recognized to be accepted, and once recognized, whether to serve it remains the other side’s decision — the same wall the same-day Wallet article describes as both sides of agentic commerce building separately.

Sources and notes

Fact boundaries: the announcement is Wang’s November 12 post, read through TechCrunch’s November 12 report in full (November 14); Meta’s product and subscription pages offer no publicly readable calling or pricing details, so the early-adopter history, prices and competitor references all follow the report. This site registered nothing, tried nothing and placed no call; no reliable public figures exist for call or task completion rates.

18Gemini · The outreach line widens again中文

Agent calls drop to a free tier — with bills, groceries and payments attached

Editorial diagram: three columns compare three errand routes — Gemini Tasks·Google (free from Nov 2026 across US/JP/KR/CA/AU/IN: business calls, Gmail bill analysis, checkout via Google Pay, RCS add-ons to Instacart, forwarded email as calendar events; runs in the Agent Compute Engine hardware-isolated sandbox per account, confirm before buying with a spending cap; every call opens as disclosed AI and merchants can opt out anytime); Muse·Meta (from $20/month, open to paid subscribers Nov 12: the calls you avoid, shopping options and discounts, plans and checklists; runs on Muse Spark with a personal Secure VM; many hang up once they hear AI, as TechCrunch reports); Xiaowei·WeChat (free, test scope expanding: everyday threads across chat, Moments, official accounts and Channels, built on payments and mini-programs; WeLM custom model, 80B live and 617B in development, some inference expected on-device; 30M DAU inside WeChat's 1.449B monthly users). Bottom timeline, from Duplex to Tasks across roughly a decade and three generations — 2018 Duplex launches (closed-domain RNNs), 2022 LaMDA enters live calls (hybrid stacks), 2025–26 agentic calling in Search and Maps, 2026-11 Tasks goes free (unified on Gemini, six countries, bills, AP2 payments, RCS).
An editorial diagram from Google's official blog (Nov 17), the Google Research retrospective (Nov 20) and Tencent's earnings disclosure (via IT之家, Nov 18): what each route costs, where it runs and who picks up, plus a decade of Duplex lineage. It is our editorial explanation, not a product interface or call record; figures and boundaries follow the text and sources.BitShovel editorial diagram · based on official blogs and public reporting · Original source ↗View full image ↗

On November 17, Google made Gemini's Tasks free: the channel that calls businesses for you — booking a table, scheduling a service — appeared in May with Gemini Spark, then sat behind the AI Ultra tier; it now ships free with Gemini 4 Hydrogen to hundreds of millions of users in the U.S., Japan, Korea, Canada, Australia and India. The errands it takes on grew into a daily-life line: analyzing utility bills in Gmail (the official example spots a heat pump running at peak pricing), comparing and buying goods (checkout through Google Pay, receipts into Google Wallet, which opened on iOS the same day), adding items to an Instacart order over RCS texting (a photo works; a U.S. number on a supported carrier is required), and forwarding email to gemini@google.com to pull dates out as calendar invites. Calling itself is available in every U.S. state except Indiana, Louisiana, Minnesota, Montana and Nebraska.

The execution layer deserves as clear a look. Tasks runs on Agent Compute Engine: one hardware-isolated sandboxed virtual machine per account, with prompt-injection filters and Safe Browsing; purchases ask for confirmation first, users can set a hard spending cap, and on some sites the final click returns to the person. Payments ride the Agent Payments Protocol (AP2), where Google says it is working with Mastercard and Visa, with American Express coming soon. The Wallet piece split agent shopping into “can it transact” and “who vouches”; Google's answer this time is to lay its own track and plug the card networks in directly.

The same week, Google Research looked back: this channel is not a new story — it is Duplex, introduced nearly a decade ago. The architecture has had three generations: before transformers, bespoke closed-domain recurrent networks trained per task; in 2022 LaMDA entered live calls inside hybrid stacks that paired generative phrasing with deterministic state tracking and policy guardrails; now the stack is unified on Gemini, where a task is defined by free text instead of a vertical, so multi-constraint requests — schedule car maintenance around calendar conflicts — are handled in one pass. The receiving-end question from the previous chapter has been part of Google's design since Duplex: every call opens by disclosing it is an AI calling from Google, merchants can opt out in their Business Profile or simply say “please stop calling my business” mid-call; and Google adds a scale figure — business details verified by Duplex calls have been displayed in Search and Maps over a trillion times (as shared in 2023).

China's route differs — and so does its scale. On November 18, Tencent disclosed that WeChat's personal AI assistant “Xiaowei” has expanded its test scope to 30 million daily active users, woven into chat, Moments, official accounts and Channels; behind it is WeLM, a custom model tuned for cost (the 80B version is in use, a 617B version in development), with some inference expected to move on-device over time; its agent capabilities ride WeChat's payments and mini-program ecosystem and the connections to millions of merchants. Side by side, the division of labor is clear: Muse treats outreach as a subscriber privilege; Google dropped it to a free default and can explain from a decade of architecture why live calls are hard (sub-second latency, noisy rooms, reading polite refusals); Xiaowei does not dial out — it takes errands inside chat through the ecosystem. The scale references differ too: Muse took 40 days to reach 10 million U.S. downloads, while Xiaowei stands inside WeChat's 1.449 billion monthly users (Tencent's Q3 figure). This site registered nothing, placed no call and verified no order.

Sources and notes

Fact boundaries: Google's official blog (Woodward, Nov 17) and the Google Research retrospective (Nov 20) were both read in full by this site (evenings of Nov 20 and Nov 21); the six countries, the five excluded states, AP2 partners, RCS conditions and sandbox details follow the official wording. Xiaowei's DAU, WeLM and the on-device plan are Tencent earnings disclosures read through IT之家's November 18 report in full (Nov 21); the “10 million downloads in 40 days” figure carries over from the November 14 chapter. TechCrunch's November 17 brief was the entry lead (read in full November 18). This site registered nothing, placed no call and verified nothing.

19Firsthand accounts make the change concrete中文

How far did shopping, project handoffs and outreach actually get?

A TechRadar writer asked Muse to find shorts: it requested a size, returned options and a discount, then asked for an Amazon login before purchase. In a separate savings challenge, the writer reported about $35 saved against a promoted $1,000 target. The accounts show steps from options toward action, while keeping one person’s result distinct from general performance.

dots’ project accounts and Grok Bot’s outreach job raise another question: once work resumes, does it return somewhere the person can inspect and continue? The former has completed summaries and crawls alongside unfinished email and reporting; the latter moved invisible files into chat. The concrete observations are completed steps, usable outputs and who takes over next.

Personal Agent Bench adds external teardowns and task records for Muse and Instinct. Its first-round results use different task sets and sample sizes, so they are not a head-to-head ranking; the team also discloses that it builds agent software. The material helps ask better questions about runtime, memory and outcomes, ahead of future matched-task comparisons.

Sources and notes

BitShovel organized official materials and these public accounts without testing all four on one task. Personal Agent Bench’s September 22 first-round report gives Muse 19% of 90 tasks (95% interval 13–25%) and Instinct 43% of 159 (36–50%). Different tasks and incompletely published batches make them separate records rather than a ranking.

20Connect it to your own work中文

Use a familiar job to judge whether an ongoing handoff helps

Start with a familiar job whose finish line you know. Interview and social-material revisions explain dots’ creative continuity; family travel, school notices and supplies explain Muse’s everyday outputs; weekly research, drafts and recaps explain Grok Bot’s ongoing role; coordinating rides or repairs by message explains Instinct’s interface. These scenarios help frame a choice, then your own results test it.

Specify the output: a checklist, review-ready drafts or changes in the original tool. Watch whether accepted requirements survive, how new material changes the result, and how it returns when blocked. Keep a clear confirmation point for payment, bookings and external sending. Inspecting results and correcting memory should be as easy as handing off work.

This topic will continue with artifacts, longer use and matched-task comparisons: revisions across dots’ creative materials, continuity through Muse glasses and Charm, visible outputs from Grok Bot routines, and completed everyday arrangements with Instinct. Those concrete changes will develop the relationships between these four products.

What we are following

21What comes next中文

What we are following

  • What actually completes, and what needs a person, when dots maintains recurring tasks, interview materials and cross-device handoffs?
  • When will Charm shipping, price and battery details be confirmed? Can glasses and pocket hardware continue the same task?
  • Which Grok Bot artifacts, team-sharing steps and long tasks can be inspected under matched conditions?
  • When will Instinct airport rides, handyman bookings and follow-ups have a visible real-task account?
  • When can the four personal agents be compared using the same public input, versions and connections?