与TodayAl聊个人agent：它就是”the next big thing’
Chatting with TodayAl about personal agents: It’s the “next big thing”
Summary
What personal AI should truly answer is: how can we lead more meaningful lives? By Cheng Manqi; edited by Song Wei. In the past month, personal agents have become the hottest new direction for AI products. Today AI is
What personal AI should truly answer is: How can we lead a more meaningful life?
Author丨Cheng Manqi
Editor丨Song Wei
In the past month, personal agents have become the hottest new direction for AI products.
Today AI is the Chinese company that first launched a universal personal agent product, with the global version going live on September 8 and the China version on September 24.
The company was founded by Qi Junyuan. He started his entrepreneurial journey in college and launched the project‑collaboration product Teambition in 2012. After his first startup was acquired, he joined Alibaba and ByteDance, working on Alibaba Cloud Drive, the AI revamp of Feishu, and the desktop and web versions of Doubao. He left in late 2025 to start a new venture. Today AI has so far completed several financing rounds totaling hundreds of millions of dollars.
We talked with the core team about market‑focused questions: Why are personal agents making a comeback now? Will this be another fleeting trend? What are the key competitive factors?
They shared some of their current core judgments:
Being able to proactively provide services is the most important trait of a personal agent; and the prerequisite for proactivity is personalization.
A personal agent is a systems engineering effort, with at least five key competitive factors: model, personalization, Infra, product definition, and customer acquisition capability.
In internal evaluations at Today AI, some new models are getting stronger at coding, but their interaction and natural language abilities are deteriorating. The optimization goals at the model layer have diverged, which is one opportunity for new personal‑agent companies.
The underlying assumption of the above judgment is that a personal agent’s long‑term role is not only as a “assistant” but also as a life partner. Making restaurant reservations or buying things for someone is only a small part; its real value lies in eventually helping us understand who we are and what kind of life we want to lead.
Most people can’t achieve this now, partly because they’re trapped by countless trivial tasks and daily frictions. An AI that spends a long time with you can not only free us from chores but also give timely reminders: an overlooked health issue, an important work or family decision, a concert you missed and will regret, or a once‑in‑a‑century astronomical phenomenon.
“With my current workload, my HRV actually increased by 10 ms compared to six months ago.” HRV is “heart‑rate variability”; a rise means the body is in a relaxed, recovering state with stronger adaptability. As heavy users, the lives of Today AI team members are already changing.
This is the new motivational philosophy of the AI era, or does it truly point to a better way of living? Most people believe they need a better life and are willing to pay for it. That’s why many start experimenting with personal agents. The question is, how many will stick around.
This market has hype, but it is still “the next biggest thing”
Later: Why has the personal agent space suddenly become so lively?
Today AI: The fundamental reason is that it is the next big thing. AI ultimately needs to answer: Is it only able to help a few people accomplish limited tasks, or can it truly help many more solve numerous problems we can’t even imagine today.
Teams that have already launched products officially have been investing months ago; this is not a product that can be rushed out in a few days. Of course, after the first batch of products is released, others will copy faster.
Later: You started developing Today AI at the beginning of this year. What inspired this product form?
Today AI: Actually, I was already envisioning a similar product back when I was building an AI browser in 2023.
The pattern I observed is: a group of people have been imagining whether there could be an AI system that’s always there to help me handle various tasks. After coding agents became more usable a couple of years ago, they started using long‑thread tasks for these experiments, and it gradually spread among pioneering users in the tech community.
Later: Why are more people working on this direction this year?
Today AI: The OpenClaw release earlier this year was a milestone. It attracted many non‑AI professionals; although it was quite difficult to use and most people eventually gave up, the fact that “regular users tried it” really happened.
Later: How can we determine whether this wave of personal AI will truly reach ordinary users, rather than fading like some past personal AI assistant products?
Today AI: Today we have indeed found a way to turn data into intelligence. I remember that early on, Siri used a decision tree to provide Q&A capabilities, which couldn’t exhaustively cover the myriad needs in people’s lives, so it couldn’t scale. On this point, I remain confident.
Later: Will there be short-term setbacks? Currently, the two hottest products have unique growth strategies: Meta has heavily promoted Muse, while Instinct runs a hard-to-get invitation testing program.
Today AI: That’s exactly my concern—the maturity and commercial value of personal AI assistant products today don’t match their valuations and hype. This direction may experience significant volatility and hype.
Later: Amid the hype, what adjustments are you making?
Today AI: The biggest change is that the domestic version was launched early. Many people in China will chase this hype soon. So we need to quickly tell everyone: a company has been doing this earlier, and the product is already usable. We are the leader of the big battle, not just joining the crowd or following the trend.
Later: Companies like ByteDance, Alibaba, Tencent and other large domestic tech firms are indeed brewing personal agent products. Who do you think can do it better?
Today AI: I really can’t see which of these three is significantly better or at a clear disadvantage. The key is that personal AI assistants are just starting, and our thinking about the core competitive factors may still be incomplete, so it’s hard to judge who will win right now.
Later: What important competitive factors have you seen so far?
Today AI: models, personalization, Infra, product definition, customer acquisition. Another factor is hardware, but hardware’s penetration speed is limited, so I’m not sure whether hardware will become the decisive factor.
Later: What is the priority order of these elements?
Today AI: It’s not about ranking; if any element is missing, we shouldn’t do it.
Later: So you also train models yourselves?
Today AI: From day one, we planned to build our own models. Because models drive most of the product’s processes and outputs, if you want to define a different or better experience, you need to train the corresponding capabilities into the model.
Later: Does doing it yourself refer to post‑training?
Today AI: Post‑training, mid‑training, and pre‑training are just methods. What matters is whether you have a model that serves your product goals. If pre‑training is needed in the future, you can also obtain the necessary resources to do it.
Recently we have started pre‑training models, and we already have an early experimental version.
Later: What was the original plan?
Today AI: First software, then hardware, then model. Now it has become doing software and model simultaneously, and then considering hardware.
Later: Will the hardware be built in‑house as well?
Today AI: We will work together with partners. There are two core assumptions behind this: first, a personal AI assistant needs more diverse hardware to obtain richer personalized context, such as health data from rings and wristbands; second, the hardware itself will become increasingly varied and personalized—recording devices can be headphones, wristbands, magnetic phone backs, etc.—different form factors make sense for different users.
From the user’s perspective, the future better experience will be a unified brain at the software/model layer handling everything, while hardware forms and brands will be abundant. So the end result must be an open‑collaboration model.
Later: What happened that made you decide to train the model ahead of schedule?
Today AI: During product development we found that some experiences can no longer be achieved without changing the model. A simple example: recent generations of models have become increasingly strong at coding, but their interaction quality has actually worsened, becoming more “unhuman”. When you ask them to write something or explain a concept, the experience may be worse than before. In our internal evaluation set for personal AI assistants, scores for some new models are also dropping, which is an important signal that everyone is optimizing in different directions.
Later: If OpenAI and Anthropic both struggle to balance higher intelligence with “speaking human language” during interaction, why can you do it well?
Today AI: It’s not just a capability issue, but a goal issue. They may focus more on optimizing other abilities, while we place greater emphasis on interaction ability.
Later: What constitutes good interaction ability? How can it be broken down?
Today AI: I think the most important feature is personalization. This isn’t something current models naturally have; it requires dedicated design and training. As use cases keep expanding, the required capabilities will increase, making this a continuously evolving direction.
Later: Why didn’t you list compute resources as a competitive factor?
Today AI: Ultimately, this is a problem that can be solved with market and resources. Startups can also obtain more resources.
Later: What might be the shortcomings of large companies building personal agents?
Today AI: They didn’t start just because they wanted to do this; they saw a business opportunity. And many of our product decisions are because we want to do something, not because we copied anyone.
Later: Can’t imitation be an efficient way to succeed?
Today AI: This field is in its early stage; the product forms we see today may be completely different from what they should look like in two years. There will be many concrete product and technology choices ahead. At that point, understanding what truly drives you becomes crucial.
Later: One option is to go more vertical, focusing on specific scenarios such as travel, health, or family affairs—Google, for example, launched the family‑collaboration agent product CC. It could also target vertical user groups, like high‑spending consumers who pursue a high quality of life. But this is not Today AI’s choice, why?
Today AI: This can only create some differences in positioning and interaction experience; but in core capabilities, it’s hard to form a real distinction from generic solutions, so I think it’s not very necessary.
Later: After the product form matures, what will differentiate personal agents?
Today AI: There won’t be much differentiation; ultimately, a limited number of companies will provide similar offerings.
Later: Will the supply become more concentrated?
Today AI: It may be like telecom operators and banks, where a few companies dominate a large market.
Later: How can we become one of those few companies?
Today AI: Continuously provide each user with the best possible experience for them.
Later: If the offerings of the few companies are similar, what are the barriers for these products?
Today AI: A personal AI assistant naturally incorporates data accumulated from a user’s long‑term usage, making the agent increasingly tailored to that person. After enough time, such products become hard to replace. So I believe that when the basic capabilities are similar, the superior experience built over long‑term use and the switching cost are the main barriers.
What personal AI should truly answer: How can we lead more meaningful lives?
Later: OpenAI calls Dots an always-on agent, Meta calls Muse a personal agent, Today AI and Instinct both refer to it as a personal assistant. What different product thinking lies behind these word choices?
Today AI: The difference between agent and assistant is that the former is relatively unfamiliar to most ordinary users. Meta may be quite confident, so it directly uses agent. Dots does not focus on personalization, but emphasizes “always on” and long‑term conversation.
Later: What does “personalized” mean?
Today AI: We need to distinguish two terms here, one is personalization, the other is individuation. Personalization refers to the agent entering non‑productivity scenarios. Instinct, Muse, and Today AI are clearly different from previous products in this regard. As for Dots, based on my own experience, it is still more productivity‑focused, and I find it hard to put personal life information into it.
Individuation, on the other hand, means whether an AI system truly strives to provide each person with distinct assistance, respecting individual preferences, the shared experiences between it and the user, and being able to discern the user’s deeper motivations. I think individuation is not yet a consensus today.
Later: At what level of individuation are different products such as Muse, Instinct, Today AI, etc., currently?
Today AI: It’s still too early; I find it hard to draw that spectrum. I also don’t know what other companies truly believe or are doing.
I can only say we strongly believe in individuation. The reasoning behind this is that the eventual widespread adoption of AI heavily depends on whether it can proactively help users discover and solve problems. And individuation is the premise of proactivity.
Later: Why is proactivity so important?
Today AI: Actually, I think proactivity is the most essential trait and source of value for a personal AI assistant. Everything else—personalization, model capabilities, infrastructure cost optimization, engineering prowess, etc.—is just paving the way for proactivity.
If it lacks initiative, it can’t help you avoid missing out on important things: those regrets about things you didn’t do, which could have turned out differently if someone had reminded you at the time. This could be a crucial health, family, or work decision, or a rare event, exhibition, or sports competition.
The premise for AI to accurately remind you of these matters is that it knows you well enough. It understands what truly matters to you and knows when to bring it up.
Later: Your description of the value of a personal agent or personal AI assistant doesn’t point to a purely tool-like value—what is the broader picture behind it?
Today AI: The core is, I think that someday in the future, at least for ourselves, we want to better understand what kind of person we are.
Many people can’t achieve this now because a lot of their energy is spent on various frictions and trivial matters. Then, three to five years later, there will be a moment when you think: What have I been doing these past few years? I’ve missed many possibilities in life.
The future won’t be like that; AI will help us handle most of the trivial tasks and frictions, allowing us to spend more time on truly important things. Those things you previously didn’t do, or even didn’t think possible, because of temporary cognitive or skill bottlenecks, will happen more often.
At the same time, the range of tasks AI can help people with will keep expanding, requiring increasingly broad capabilities. So this direction is very long-term; we still have many gaps to fill, one by one.
Later: How did you arrive at this level of value?
Today AI: It may be a realization I’ve had after years of entrepreneurship. It also has a prerequisite: you need to appreciate the diversity of people, society, and the world. Otherwise, the motivation behind your decisions and persistence in something will be very different.
Right now, everyone is scrambling to pursue short‑term paths that have been validated in stages, such as the industry’s focus on how a personal AI assistant can help people make transactions—ordering takeout, booking restaurants, even discussing its network effects and calculating GMV numbers. From a business and capability‑stage perspective, these are also correct. But if you think this is what personal AI is, I find it still too limited. It is far, far away from truly changing how people live, work, learn, and experience life. So I’m also reluctant to define a winning move for this direction so early.
Later: What we see now is an assistant, whereas the future you describe is a “life mentor.”
Today AI: Think about it—large models have compressed almost all of humanity’s digitized knowledge during pre‑training. At the same time, their information‑processing bandwidth far exceeds that of a human brain that runs on about 20 watts. So in terms of knowledge breadth and information‑processing capability, they can provide each person with far more assistance.
We should let AI help people experience a more meaningful, richer, and more exciting life—that is the long‑term question a personal AI should answer.
Some uninstall in two minutes, others can’t live without it
Later: Looking at the current product form of personal AI assistants, what proven value and design aspects exist?
Today AI: First, it will have a long‑term memory system, and the user’s conversations with it occur within a continuous dialogue: you no longer need to start a new conversation for each task, return to a historical thread, or think about choosing the right workspace.
From left to right, they are the user interfaces of Today AI, Instinct, and Muse respectively. Instinct’s main interaction interface is the user’s iMessage; it does not have a standalone app. Muse and Today AI have applications across mobile, desktop, and other platforms.
Long‑term conversations + memory archives have become the mainstream interaction mode for personal AI assistant products, whereas earlier ChatBots and office agent products were more about opening a separate conversation for each task.
Today agents can also connect to multiple of your information sources, proactively reminding you of insights you might miss. They can operate browsers and computers, handling tasks that users find cumbersome.
And there is now some consensus on the technical architecture: give agents their own execution space in the cloud. This way, users can launch complex tasks from different devices such as phones.
Later: After Today AI’s official public beta for a month, what things have become clearer to you?
Today AI: First, it confirmed that personalization is crucial—features that are completely useless or even terrible for one user can be the most important for another.
Later: Can you give an example?
Today AI: A typical example is that Today’s assistant has its own profile page, which includes who it is and the records left after daily interactions with the user. There’s a little Easter egg on the profile page—if the user scratches off a hidden area like a lottery ticket, they’ll see a golden phrase or blessing the agent gave them that day.
Many users think this isn’t useful, but some users love it; they draw great support and even crucial inspiration from that sentence.
There are many similar variations. For example, what conversational style users prefer, how detailed a task should be completed, which topics the agent should proactively discuss, how to initiate communication, and whether to keep everything in one conversation or use separate, dedicated threads for different matters… So I believe a good solution is personalization.
Additionally, I’ve found that the user base and behaviors are very diverse, indicating that this product form can reach a broader audience than we imagined, and they’ve created many use cases we never anticipated. Therefore, we will focus on cost optimization next. After reducing costs, we can also run more proactive strategies. I still believe that proactively offering services will create huge value.
Later: Some users mentioned that after trying Today AI, they start delegating more and more daily tasks to it, while other users uninstalled the mobile app after two minutes because they weren’t used to the new experience of a personal AI assistant you mentioned earlier—e.g., all interactions happen in one long conversation, without needing to switch dialogs for different tasks. This actually makes people accustomed to ChatGPT feel uncomfortable.
Today AI: That’s also part of the different preferences I mentioned earlier. So we also launched a multi‑task system, but the vast majority of users still interact through the main conversation. Today will proactively identify complex tasks, move them to a separate task, and sync subsequent messages back to the main conversation. If a user wants to continue issuing commands for a specific task, they can enter that dedicated task to interact, though relatively few people use it that way.
But the phenomenon you described does exist. Today has a very clear characteristic: users are either highly sticky or can’t get it to work and give up quickly, clustering at both ends. Currently, the former group makes up a larger share.
Later: How’s the overall retention after a month of public testing? Are you satisfied?
Today AI: Today has the best retention among all the products I’ve built over the years. Moreover, its retention curve shows a clear tail‑up, entering a state where it no longer continuously declines. Some users may use it a bit less on certain days, but when something new comes up they return, and usage frequency quickly recovers—there’s strong inertia.
Next, we will explore this topic: how to guide a new user to quickly find an experience of doing something together with an agent.
Later: In a sense, polarized user behavior can also be considered good feedback.
Today AI: Yes, everyone being equally bad, or everyone being equally “okay,” is the worst outcome.
From Exhaustion to Intentional Living
Later: What do you think is a reasonable business model for personal AI assistants in the future? Today AI currently has a free tier; after reaching a certain usage limit, users can choose a $20 or $200 per month subscription plan. Muse follows a similar free‑plus‑subscription model, while Instinct is completely free at this stage.
Today AI: First, let me clarify that our subscriptions are targeted overseas, and the domestic version is not planned to be paid at this time.
I think the current business models are inadequate because they don’t fully align with users’ goals. For example, charging per token encourages the platform to have users consume more tokens, while users want to solve problems using as few tokens as possible.
I would prefer to earn money by providing something users need with the highest efficiency available in the market; even after adding a reasonable service fee, the result is still better than any alternative. In that case, our goals align with the users’.
Later: The service fee you mentioned applies when you help a user complete a specific task. Earlier you also described another “mentor‑like” value of a personal AI assistant, which proactively reminds users not to miss insights, events, and possibilities. How would this part be commercialized?
Today AI: I haven’t figured it out either. Because its revenue is highly personalized, it might eventually become a basic service fee.
Later: It seems this would require integration with existing e‑commerce, local‑life platforms, and apps, and they may not be willing to open up—for example, Amazon just banned Muse.
Today AI: At this stage, we’ll do our utmost to negotiate partnerships one by one. Ecosystems aren’t static. When new technologies, experiences, and user value objectively exist and the old ecosystem can’t adapt or meet the new demands, new solutions will emerge. Anyone could build them, including us, so there’s no need to rush.
Later: What do you think a healthier business ecosystem looks like for personal AI assistants?
Today AI: I believe any platform advantage built on user habits, laziness, and ignorance will be broken. With agents, information asymmetry will decrease, and users will no longer need to do a lot of comparison and operation themselves. In the end, it comes down to who can truly deliver value to society and who can provide the best products or services for a particular group of users.
Later: Specifically, which platforms or services might be affected?
Today AI: For example, ride‑hailing—whether Didi, Amap, or Cao Cao—when I’m about to go home, there should theoretically be a most cost‑effective option, but most people won’t operate three platforms simultaneously. E‑commerce is the same; it’s hard for me to compare Taobao, JD.com, Pinduoduo, and even self‑operated versus third‑party items on the same platform at the same time. Sometimes, when we can’t judge which product suits us best, we end up choosing the one with the most ads, the most recognizable brand, the one that looks least likely to go wrong.
At that point, if a personal AI assistant can fully understand my specific situation and preferences, and also compare different platforms and products, it can help me decide what to choose. The extra revenue that platforms previously earned due to user inconvenience and incomplete information would be diminished.
Later: According to your reasoning, will bilateral network effects still exist in the future?
Today AI: There will be huge changes. For example, a creator has their own shop, selling the items that are most suitable for a certain type of user within a category. Users can have their own agent transact directly with them; neither side needs to pay an intermediary information platform, and the whole process can even operate without a fixed software platform or interface.
Later: But heavy aspects such as logistics and offline services should still exist.
Today AI: We need to separate the different values of the platform. Logistics and other fulfillment have real service costs, so fulfillment capability still has value. Only if AI or robots can also replace part of human costs, creating new productivity, will the business structure change further. But today many platforms gain not only fulfillment revenue, but also a large amount of information intermediary revenue, and this part of value will be hit hard. Fulfillment itself may also become more decentralized, because agents can reorganize dispersed service supply.
I personally have a clear feeling now: I increasingly don’t need review and navigation sites. Because an agent can, at a given moment, provide sufficiently good results based on my personalized context. Product forms and service models that once seemed completely reasonable no longer serve me that well.
Personal AI assistants are reshaping information intermediaries, potentially reducing profit from asymmetric information between consumers and service/product providers in the matchmaking stage.
Later: When do you think more ordinary users will start feeling this way?
Today AI: With today’s models and product maturity, I’m not confident it can scale massively within a year. It will likely be a gradually accelerating process.
First of all, I am a very early‑adopter user, love trying all kinds of uses, and can tolerate failure better, but such user motivation is certainly a minority.
Second, from the perspective of the system actively understanding and adapting to users, our proactivity today is still only about twenty to thirty minutes. If we still have to rely mainly on users learning how to use it themselves, the adoption speed will naturally be slow; this is one of the reasons why proactivity matters.
Later: As a frequent user, how has Today AI changed your everyday life?
Today AI: Previously I had to spend time figuring out what people were discussing in emails, calendars, Slack, and other collaboration tools; now that time has been greatly reduced. This allows me to shift from many organizational and trivial tasks to more important issues.
But I’m also correcting something. I love being hands‑on, and with more time, I sometimes over‑execute tasks that should actually be delegated to the team.
There are also health changes. Under my current workload, my HRV has actually increased by 10 ms compared to six months ago, which isn’t easy to achieve.
Today also improved the quality of my social interactions. It helps me prepare for interviews, meetings with investors and partners, and my communication quality has increased a lot.
Later: You started using the early version after March. When did you noticeably feel these changes?
Today AI: It was around the April‑May version. At that time, the team wanted me to migrate my account from an external environment to an internal one, but it never succeeded. Whenever I felt the new agent had reached about seventy‑or‑eighty percent maturity, and I occasionally mishandled something, I would switch back to the old agent because the old agent understood my preferences better— even a slight edge makes a huge difference in experience. This is what I mentioned earlier: the switching cost for a personal AI assistant is also high.
Later: If more people, like you, let personal AI assistants handle trivial tasks and friction, what could they devote a lot of time to? What would life be like then?
Today AI: A lot. People’s lives mainly consist of rest, work, socializing, and hobbies, and each of these areas has a significant chance of improvement: you might experience a more enjoyable work process, discover new job opportunities, spot previously overlooked health issues early, and take timely care of family and friends…
A person who is constantly rushing versus someone who lives intentionally have vastly different physical, mental, and life states, and most people I see now fall into the former category. This is an area where everyone can improve.
Image source: generated by GPT. The original image is Wilson, Chuck Noland’s companion in “Cast Away.”
– FIN –
Get MirrorSpeed