Personal AI Assistants Win by Holding Your Whole Life

Meta’s Muse reached No. 1 on Apple’s U.S. free iPhone chart on September 18, ten days after launch. Instinct, a text-first personal AI assistant, is in talks to raise $1 billion at about $10 billion.

These products work by staying in mail, calendars, browsers, and phone lines, then acting after the app is closed. That stay is the feature investors are buying, and it is the part users cannot take with them.

Instinct Is in Talks at $10 Billion

Spear Street Technology, the San Francisco firm behind Instinct, is negotiating a round of about $1 billion, a person with knowledge of the deal said. Sequoia Capital and Benchmark have been in talks to lead. Chief executive Noah Shinn, 23, had no comment, and the terms are not final.

The figure would be four times the $2.5 billion mark from August 26, when Index Ventures and Benchmark led a $250 million Series B that took total funding to $350 million. Earlier in the spring the company had been valued at about $50 million. Backers also include Conviction Partners and Greenoaks.

THE FUNDING SPRINT

1. Spring 2026: Spear Street is registered and Instinct is valued at about $50 million.
2. August 26, 2026: A $250 million Series B prices the firm at $2.5 billion, bringing total capital to $350 million.
3. September 16, 2026: Talks surface for about $1 billion at about $10 billion, with Sequoia and Benchmark named as possible leads.

The raise is being shopped while the product is still free. Shinn has said he does not want to charge people, and has pointed to advertising as one way the company might pay for the work. Users text Instinct to answer mail, negotiate bills, and book tables. The invitation-only service has passed more than 100,000 users and has told some of them it is at full capacity, with slower replies.

On September 16 Shinn said Instinct Concierge can place phone calls for early-access users, a white-glove path for restaurants that do not take bookings online, dentist cancellation lists, and cable-bill fights. The same week the agent got its own email addresses so it can open and run accounts directly.

Muse Took the U.S. App Store

Meta opened Muse to U.S. users on September 8 as an iPhone, Android, and web agent that also lives in WhatsApp. Sensor Tower counted about 730,000 U.S. downloads after launch. A Mac build arrived on September 17, and outbound calling to U.S. businesses landed on September 16, the same day as Instinct Concierge.

MUSE IN ITS FIRST DAYS

• Chart: No. 1 on Apple’s U.S. free iPhone list on September 18, ahead of ChatGPT, Gemini, and Claude.
• Access: U.S. iOS, Android, muse.ai, and WhatsApp, with AI glasses listed as coming later.
• Meter: A free weekly token cap of about 100 million, with iOS paid plans listed at $20 and $100 a month.
• Checkout: Stripe Link mints a one-time card so Muse never sees the real number.

Meta’s newsroom says each person gets a dedicated secure virtual machine with its own browser, and a separate Sentinel process that must clear anything headed to the internet. Muse is not supposed to see passwords or payment methods. People pick which apps it can read or send from, can tell it to forget a detail, and can opt out of using chats to train Meta’s models. Meta also says Muse does not share those conversations or the VM’s data with its ad systems.

When it is time to pay, Muse can check out through Link by Stripe and is the first agent covered by Link purchase protections for agents, including damaged or lost items, price drops, and no-fee returns on eligible buys. A Confidential VM, encrypted with a key only the user holds, is promised later in 2026, so even Meta could not open the box.

Chief AI officer Alexandr Wang called the App Store ranking a week after launch. Mark Zuckerberg has described Muse as a system that works around the clock on goals, health, relationships, and money. The pitch is control. The design still needs mail, calendars, Instagram saves, OpenTable, Ticketmaster, Shopify, and a card on file before it can do the interesting work.

Five Agents and What They Can See

The consumer field compressed into a few weeks. Grok Bot, Instinct, Muse, and Siri AI all shipped or relaunched between August 11 and September 14, while OpenAI is still converting a February hire into a product people can text like a concierge.

WHERE EACH AGENT LIVES

Google still has Gmail, Calendar, and search, and it did not put a life-running agent into this September burst. Anthropic’s Claude Cowork is on the desktop for people who already live in that stack. The fight that drew the new money is the consumer inbox, not another chatbot window.

Why Apple Put Caps on Siri AI

Apple began rolling out Siri AI on September 14, 2026, with the 2027 software releases, in an English beta. French, Japanese, Korean, Portuguese, and Spanish are due in October. A new Siri app syncs chat history through iCloud across iPhone, Mac, iPad, Apple Watch, and Apple Vision Pro. Apple says the rebuilt assistant can search messages, mail, and photos, act on what is on screen, and draft mail in the user’s usual tone.

It can add a relative’s recipe to Reminders after finding the suggestion in a thread, or pull home games off a team site and drop them on the calendar. That is closer to a secretary than the old voice box. It is also the work that leaves the device.

Certain Apple Intelligence features that rely on server-side models are subject to daily usage limits, including but not limited to Siri AI, intelligent photo editing tools, Image Playground, and AFM 3 Cloud models in Shortcuts. Daily limits may vary by feature, request complexity, system demand, system policies, and other factors. Expanded access to such features will be available for a fee in the future.

Apple, Siri AI newsroom note, September 14, 2026

Those daily usage limits on Siri AI sit in the footnotes, not the demo. Apple has not published a number, a start date, or a price. Private Cloud Compute is still the privacy claim: personal data is not stored or made available to Apple when a request runs in the cloud, and outside researchers can keep checking. Siri AI is held back in the EU on iPhone, iPad, and Watch, and the new Apple Intelligence features are not shipping in China while Apple works through rules.

The on-device path is the part that does not need a new relationship with a startup. The cloud path is the part that will, eventually, look like a plan.

OpenAI Hired OpenClaw’s Builder in February

On February 15 Sam Altman wrote that Peter Steinberger was joining OpenAI “to drive the next generation of personal agents.” He called Steinberger a genius with ideas about smart agents working with each other, and said the work would become core to OpenAI’s products. Steinberger had built OpenClaw, the open-source messenger agent that let power users run mail, shopping, and calendars from WhatsApp and similar chats. His own line for the next job was to build an agent his mum could use.

Seven months later that consumer product is still the missing piece. OpenClaw moved to a foundation OpenAI said it would support, and the OpenClaw shared workplace update pushed the project from a solo bot toward sessions a team can hand off. Inside OpenAI the agent stack went to work instead: ChatGPT Work for desk jobs, an Agents API for cloud harnesses, and, in September, tests of Sponsored Agents, brand-tuned chats that start from ads in ChatGPT.

The company already ships GPT-6 Astra computer-use tools that can operate software. What it has not shipped is a Muse-style proxy with a phone number, a standing login to Gmail, and a brief to run a household. The hire bought the person who proved people would give an agent those keys. The September products that actually hold the keys belong to Meta, Instinct, SpaceXAI, and Apple.

The Dinner Reservation That Got Banned

SpaceXAI launched Grok Bot on August 11 as “AI teammates you can give real work to,” each with an own computer that keeps working after the laptop lid closes. Bots sign into the apps a person already uses, finish jobs end to end, and return only when something needs a yes. On August 26 the company folded that access into SuperGrok and paid Cursor plans, with a usage pool separate from ordinary Grok chats. The computers run in Cursor’s U.S. cloud. There is no on-prem option.

That always-on machine is the same idea Instinct sells over iMessage, and it collides with firms that never wrote an agent into their rules. JC Bahr-de Stefano, a principal at Better Tomorrow Ventures, texted Instinct to get a table at 4 Charles Prime Rib in New York. The activity log he pulled later showed the assistant hitting Resy on the order of 200 times an hour, including bursts around the morning release. Resy deactivated the account and canceled future bookings. His caption was “Nice going Instinct.”

WHAT THE RESY LOG SHOWED

• The ask: A one-line text for a table at 4 Charles Prime Rib.
• The pattern: Roughly 200 requests an hour, per the diner’s own log, including sub-second polling at release time.
• The result: Account off, future reservations wiped, then a reinstatement with a warning.
• The rule: American Express, which owns Resy, said unapproved third-party bots may not work the platform on their own.

An American Express spokesperson said Resy “does not currently permit unapproved third-party bots or agents to independently access or interact with the Resy platform,” and that automated activity can hurt a fair shot for other diners. Bahr-de Stefano got the account back. The warning said a repeat could cost him Resy and his Amex card. Another diner, Brian Distelburger, described a similar ban after a self-built agent “went wild” on the same restaurant and never booked a table.

Restaurant software is not the only property meeting agents hitting sites that never signed up. The useful agent has to look like a person with a login. Platforms built to stop scalpers will treat it like one. Insurance on a Stripe card does not help if the reservation company has already closed the account.

Who Keeps the Agent’s Memory

On September 18 Zuckerberg opened Muse connectors for outside developers. The offer is blunt: they bring an API, Muse brings the agent, the browser, and “the context of what the person actually wants.” People reach a service by asking, and the agent takes it from there.

Opening access for developers to build Muse connectors. You bring the API — Muse brings the agent, the browser, and the context of what the person actually wants. People reach your service just by asking for it, and their agent takes it from there.

New connectors are live… pic.twitter.com/EfXqF0xf0u

— Mark Zuckerberg (@finkd) September 18, 2026

That is a storefront. Whoever sits between the wish and the booking sees which merchants convert, which ones people will pay, and which habits are sticky enough to rank. Early testers have already dropped Instinct, Muse, Grok Bot, and coding agents into one group chat and let them divide the work. The portable version of that memory is still a pile of ordinary files. The default in the new apps is a VM the vendor hosts.

will.i.am, whose FYI.AI gives agents their own phone numbers that can text, call, and work a project with a person, said he recoils at a free agent “in exchange for extracting your very core.”

That’s nasty. We’ve seen that with social media.

will.i.am, founder, FYI.AI

Instinct is raising on a free product that may need ads. Meta walled Muse off from its ad pipes and still wants the connector layer. Apple will sell more cloud turns. OpenAI is testing brand agents inside ChatGPT. The assistant that can call the cable company is the assistant that already has the bill, the calendar, and the voiceprint of how you write to your sister.

Meta’s Confidential VM is due later in 2026. Until it ships, the agent that books the table is the one that already has the login.