The AI office hype hasn't cleared, and the personal Agent war has already begun | Breaking down Town, Instinct, Grok Bot and Muse
Instinct has fewer than 100,000 users and its product isn't officially launched, yet its valuation went from $50 million to $10 billion in half a year — in this round of personal Agent heat, capital grabbing a ticket matters more than real demand.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
100,000 users support a $10 billion valuation
Instinct only began small-scale invite testing in February this year, and by the end of August it still hadn't reached 100,000 users, yet its valuation rose from $50 million at the start of the year to $500 million in August, $2.5 billion at the end of August, and this month the new round being quoted in the market has already reached $10 billion, with Sequoia and Benchmark both in talks to lead. The founder is a 23-year-old unknown with no big-company experience, no technical patents, no industry background. Pan Luan put it bluntly: what capital is buying is neither the founder nor the product, it's the story of the next OpenAI, and it can pay the price of the next-generation personal operating system to grab a ticket.
— Pan LuanThe line between tool and person is initiative
The biggest difference between this wave of Agents and the last is that it is no longer another tool, but a person. Two differences between a person and a tool: first, long-term understanding — a tool needs the context explained again every time, while a person understands what kind of person you are like a friend or colleague; second, initiative — assigning an automation task in Manus still depends on a prompt, on you spelling out the rules and intent, and that isn't initiative. True initiative is subjective agency — like the colleague who understands your goal in doing something and proactively comes over to say, "I have a plan, should we discuss it?"
— SukiOffice Agents organize by project, personal Agents organize by person
The most core difference is the dimension along which context is built. Cloud Code builds context along a mission or project dimension, and switching to the next project means starting over; this wave of personal agents builds context centered on you as a person, knowing your profile, preferences and past history. Open Codex or Manus and the list on the left is all the tasks and projects you've sent out; open Grok Bot and beside it are AI companions you can drive together. In personal scenarios people care less about model capability and the harness layer, because individuals don't have that many complex tasks; office Agents are the ones that care a lot about task completion rate.
— SukiTown first trades email for the user's right to delegate
The four products are trying to solve the same problem, just with different approaches. Town has a mobile client, but 70% of its data traffic is on PC; it focuses on the email scenario, solving one high-frequency scenario well before expanding. It understands which emails you send and receive often, auto-archives, learns your writing style, drafts replies for you, and sends automatically after you click confirm. It goes deeper in the email scenario than any other agent, and overseas it has already achieved fairly viral spread, especially in school-related scenarios. In one sentence: use a high-frequency scenario to build user trust, and bit by bit acquire the user's ultimate right to delegate.
— SukiPoke didn't break out, and it wasn't a technical limit
Poke and Instinct are both agents that live inside IM and don't make a standalone app, but Poke didn't break out. Suki believes it wasn't a technical limit — Browser Use was already very mature by the end of last year, and Poke could absolutely have done it at the time; the difference was team strategy: Instinct went deeper in the agent direction, taking the full-delegation route, able to connect to credit cards, screen and microphone, booking flights and buying things for you — things users hadn't seen before, which produced viral spread; Poke put its energy into content push accuracy, taking the proactive-push route, with email notifications arriving even earlier than the inbox itself, but it didn't take the step of "finishing the thing for you."
— SukiA cloud Agent has to handle three environments
Mark says doing a cloud agent was a painful lesson: a local agent and a cloud agent are completely different things. To do personal AI you have to be usable anytime, anywhere — you can't be unreachable just because the computer screen is off. The cloud has to handle three environments — the user's device, the cloud service, and the agent's execution environment — and task state, task interruption and recovery, and context passing between the three are very complex systems. They started in February, and just getting the agent to run stably on the cloud took at least two months; setting aside detours, polishing a sincere piece of work takes at least three to four months.
— MarkHiring an assistant and being a middleman are two businesses
Town started out doing corporate tax filing, and after pivoting it leans more toward enterprise scenarios; the relationship is more like you paying to hire a personal assistant, which naturally filters for people willing to pay, with a typical profile of white-collar workers who are extremely busy and have no time or energy. Muse is more like a middleman, shopping guide or broker: although it has a paid subscription, the main projected revenue source is still commission from matching transactions, leaning toward mass users and blending into everyday WhatsApp chats. Two completely different logics, cutting different scenarios and filtering out different users.
— SukiFor an Agent, the saved folder may be noise
The story Zuckerberg told when launching Muse was extracting what you like to eat from ingredients saved on Instagram, generating a shopping list, adjusting the menu by dinner party size and dietary restrictions, then completing the shopping and relationship maintenance — turning years of interest profile into money, going from a discovery platform to a decision-and-transaction platform. But Suki suspects this is to save face for the free business narrative: helping users complete transactions is absolutely not among the top scenarios users want solved; users more want to solve the annoying things they don't want to do themselves. There isn't that much high-quality context in saved folders; truly high-quality context comes more from offline, IM and email, and Douyin and Instagram saves may themselves be noise to an agent.
— SukiGUI is a maze map for ordinary people
Today didn't initially want to make a standalone app, and was very channel-heavy, integrating WeChat and Telegram, with most scenarios completed inside the channel. But it quickly discovered the discoverability problem: users facing a blank dialog box had no idea what to do. So it had to step into GUI, and doing GUI requires a standalone app, and in the end it still made chat in the app. Suki's takeaway is that GUI today no longer carries the user's operation entry point the way it used to, but is more like a map — to ordinary people an agent is more like a maze hiding many treasures, and GUI gives ordinary people a map of the maze, along which you can discover the treasures.
— SukiIf 90% of effort goes to generic capability, change direction
Suki originally made the AI education product Monspeak, and when doing 2.0 during this year's Spring Festival discovered that to make a truly personalized teacher, 90% of the effort ended up going into how to make the agent run stably and have personalized memory. If 90% of the effort is going into a very generic capability, then conversely those generic Agents might casually finish and eat up the vertical scenario. So the judgment was that in the C-end track not that many players will remain in the end — it will definitely be highly concentrated, winner-take-all — and so Suki switched to joining a startup to do personal AI together.
— SukiThis wave has no network effect to lock users in
Suki believes this AI transformation is very different from the last mobile internet wave: the biggest change this wave is technological innovation, not the narrative of scaling up, burning money and building network effects. There is no network effect or scale effect that can lock users in — as long as your product is good to use, I'll switch to another product the very next day, because no two-sided network platform has formed and users have no need to stay with you. People stay on Xiaohongshu or Douyin because a two-sided network platform has formed and they can't easily migrate away. So innovative companies are actually easier to fight than in the last mobile internet wave — just hurry to innovate and win users with the best experience.
— SukiThe most valuable thing is the middle processing layer
At the end they asked where the most valuable position is in the personal agent era. Suki picks trust and habit; context is only the entry-level foundation, and ultimately it's the dependence, habit and trust formed through long-term use; if forced to choose, Suki leans toward the end closer to the user, because model makers are still chasing each other and the merchant and MCP end will naturally grow with the market. Mark thinks it's neither context nor the ecosystem layer: just getting a lot of context is completely insufficient and delivers no direct value, and memory is far more complex than imagined — something said last week may already be invalid this week, and if it stays in memory it's serious noise. The most valuable thing is the middle processing layer, a bit like Douyin's recommendation algorithm; how the execution mechanism works is each company's secret.
— MarkIn their own words · checked verbatim
It's a product with not even 100,000 users, and its valuation multiplied 200 times in half a year.
就是一个10万用户都没有的产品 半年时间的估值翻了200倍
Pan Luan0:00
What I'm buying isn't you the founder, and it isn't this product either — I need the next story, I need the next OpenAI story.
资本我买的又不是你这个创业者 我买的也不是这个产品 我需要下一个故事 我需要下一个open AI的故事
Pan Luan1:01
Then he'll proactively come to you and say, I think you want to do this thing, and I have a plan, so should the two of us discuss it?
然后他会主动过来 找你说说 我觉得你想做这个事情 然后我有方案 然后我们两个 要不要讨论一下
Suki17:18
And then helping you finish the thing — that itself is very valuable, and users start making moves to use it.
然后而且帮你把事情做完 这件事情本身是非常有价值的 用户就产生了用它的动静
Suki39:33
Actually helping users complete transactions, even recommending the thing you want to buy, is absolutely not among the top scenarios users want solved; users more want to solve the annoying things they don't want to do themselves.
其实帮用户促成交易 甚至给他推荐你想买的这个东西 绝对不在用户想解决的talk的场景里面 用户更多的是想要去解决那些烦人的 他自己不想做的一些事情
Suki1:14:07
GUI today may not be a very, very important entry point carrying user operations the way it was in the past, but is more like a map.
GUI 在今天来说 可能不是一个非常非常重要的 像过去一样 是承载了用户操作的一个入口 而是它像是一张地图
Suki1:21:13
As long as your product is good to use, I'll switch to another product the very next day.
只要你的产品好用 我第二天就立马转向 另外一个产品去了
Suki1:55:42
Figures
| Instinct users | fewer than 100,000 | 0:00 |
| Muse first-week downloads | close to 900,000 | 2:03 |
| Meta stock single-day gain | 6.5% | 2:03 |
| Town user scale | about 10,000 | 9:11 |
| Town PC data traffic share | about 70% | 30:24 |
| Time to polish a cloud agent running stably | at least two months, three to four months for a complete work | 43:39 |
| Muse subscription price | $20 per month, $100 per month for users wanting more Tokens | 1:04:56 |
| Muse projected user scale | about 100 million users, of which 1% contribute subscription revenue | 1:04:56 |
| Upper limit on China's white-collar population | no more than 200 million, and fewer than 100 million pay personal income tax | 1:10:00 |
Glossary
- Personal Agent
- An AI that organizes context around the user themselves, understands them long-term and proactively does things for them.
- harness
- The engineering layer that wraps the model and is responsible for calling tools and completing tasks.
- MCP
- An open protocol ecosystem that lets AI call external tools and data.
- Browser Use
- The ability to let AI operate web pages through the browser interface in place of a person.
- Computer Use
- The general ability to let AI directly control the graphical interface to complete tasks.
- context
- The information about you that the AI can see, including history, preferences and the current task.
How to listen
Product managers, investors and engineers watching startups in the AI application layer, especially those building or preparing to build personal assistant products and needing to judge this round's valuation and product cadence.
The 01:31 to 01:45 segment with the North American junior-year founder on the mic is mostly fundraising narrative, with information density clearly lower than the segments before and after.