The AI Assistant Battle: Entry-Point Design and Monetization Challenges of Muse, Grok Bot, and Instinct

On October 21, Anthropic released Claude Duet, joining the competition in personal AI assistants. Earlier, Meta had launched Muse, SpaceX's Grok Bot aimed to build a team of digital colleagues, OpenAI also announced its own assistant Dots, and Instinct put a personal assistant into Apple's iMessage—and just one year after founding, it completed a $1 billion Series C at a $10 billion valuation.

In China, Manus launched version 2.0, Tencent's WorkBuddy also entered office scenarios starting from organizing materials and handling spreadsheets, and Qwen and Doubao have both launched agent products.

As Agents start doing work, they also bring new computing demand. When they open browsers, run programs, and process files in virtual computers in the cloud, they all need CPU, memory, and storage support, and the stock prices of AMD and Intel have also risen over the past month.

But more tasks and more computing power do not automatically equal more profit—the business behind this is not easy to do. In this episode, we interviewed several Agent practitioners, including a Grok Bot designer, to answer why big companies and startups are all rushing to build AI assistants, and whether the entrance they are fighting for can ultimately become a profitable business.

Why Are Agent Products Emerging So Frequently Now?

If you also tinkered with OpenClaw at the beginning of the year, seeing how these Agents are used might feel very familiar: couldn't "Crayfish" already do these things?

Connecting to email, managing schedules, operating the computer, and then letting it run continuously in the background 24 hours a day, proactively providing services. The basic ideas of today's personal assistants can already be seen in OpenClaw.

Howie Xu

Chief AI and Innovation Officer of Gen

What OpenClaw represents behind it is an architecture, and it represents the maturity of the technology. So today, when we see Grok Bot, including Meta's Muse, behind them they are certainly not using OpenClaw, but they are using OpenClaw's thinking to build applications with these architectures.

However, although OpenClaw proved that the architecture can be used and the technology is mature, it never entered every household, and there are reasons for that. Its technical threshold is indeed very high, which is why some people sell courses, and some people pay others to install it on-site. From another perspective, this precisely shows the demand.

Peng Zheng

SpaceXAI Designer (Grok Bot)

People do have demand for this kind of personal assistant form that can truly complete tasks for you, but they are also somewhat deterred by its threshold.

So, from OpenClaw to Muse, it is not that a completely different set of ideas suddenly appeared. This batch of products is trying to turn one direction into a service that ordinary people can use: the computer is prepared by the platform, and you do not need to prepare a Mac mini or virtual machine for the Agent; the runtime environment is maintained by the platform, and you only need to connect the necessary accounts and tell it what to do, which is much more convenient than setting it up locally before.

Behind this, two things are happening at the same time. The first is Harness (the agent runtime framework), as well as the growth in models' ability to handle continuous actions.

Going back to the furniture-selling example, writing the introduction may only be the first step. It also has to check prices, find the correct page, fill in information, and handle buyers' replies. If things do not go according to plan, it also has to decide what to do next. These steps have dependencies; if something goes wrong earlier, all subsequent operations may be wasted.

Beyond the model, a whole system is also needed to help the Agent connect tools, preserve progress, arrange follow-up tasks, retry before errors occur, or hand tasks to others.

Homin Luo

Founder of Mapier.ai

It is like an iceberg: what you can see and the execution interacting with you is only a little bit. But behind it there is a large amount of harness, and a large amount of work and execution related to sandbox. For example, how to run many subagents is actually all running behind the scenes.

Second, some tasks can now be completed with cheaper models. This is especially important for Agents that need to run continuously.

Howie Xu

Chief AI and Innovation Officer of Gen

In addition, especially in the past six months, model capabilities and costs have both improved significantly.

Today's OpenAI Luna model, I think, is far ahead of the most advanced model from a year ago, and may even be ahead of the most advanced model from nine months ago, while its cost is one-tenth that of frontier models, roughly an order of magnitude lower, which makes it very usable.

Because we are also developing products similar to Muse or Bot ourselves, I understand consumption very well. Although it is still expensive, at least we can start discussing whether it is affordable. Nine months ago it was too expensive, and there was no need to discuss it at all.

Why is cost so critical? Because what we hand to Agents now are long-horizon tasks. Doing one thing requires repeatedly calling the model behind the scenes; while we rest, it is also constantly working. The cost accumulated this way makes economical models even more necessary.

Of course, a single call becoming cheaper does not necessarily mean the total cost of an assistant will definitely decrease. The more things it can do, the more tasks users may hand over to it.

At this point, it becomes clear why these Agent products have become usable. In essence, they are following OpenClaw's logic while building products better suited to the general public. Looking back, a wave of such products emerged after OpenClaw went viral.

At first, Anthropic launched Claude Cowork, bringing the Coding Agent's execution approach to a broader range of office tasks: reading files you authorize, organizing materials, handling spreadsheets, and delivering results you can review and modify. At the time, this also triggered a round of SaaS anxiety in the capital markets.

Tencent also launched WorkBuddy in February, exploring a similar office path. Gradually, we have seen more consumer-facing products as well. For example, Town is trying to understand your personal work patterns from the software you use every day and automate tasks.

What these products have in common is that each is connected behind the scenes to an execution environment capable of doing work. Some directly use your computer, such as WorkBuddy; some have virtual machines and work in the cloud, such as the latest Instinct, Grok Bot, and Muse; and some call local tools.

They can process files, run programs, and, after you authorize them, enter other software and services. In particular, compared with chatbots, the bigger change is that they have memory across conversations.

But what differs is how they get users to try the product for the first time, and how they get users to keep using it. Next, we will take several recently popular products as examples to look at their different designs and goals.

Three products, three ways of building long-term delegation

Instinct

Many people may not yet have an invite code for Instinct, but venture capitalists have already valued it at $10 billion.

To date, Instinct still uses an invitation system, but it already has more than 100,000 users. Most of its early users were Silicon Valley entrepreneurs and investors, and it was precisely these seasoned users' proactive promotion that attracted a wave of attention to Instinct.

Founder Noah Shinn is a 23-year-old college dropout. In 2023, he published a paper called Reflexion, studying a question highly relevant to today's Agents: if AI fails at a task, can it make fewer of the same mistakes next time?

The method they proposed was to have the Agent summarize its experience into text based on task feedback, store it in memory, and refer to it on the next attempt. This does not require retraining the model each time; instead, it improves subsequent decisions by recording and using feedback.

Later, while working as a researcher at Sierra, he also participated in research on τ-bench (the τ benchmark). This set of tests further asks: what does it actually take for an Agent to truly get something done?

For example, in a simulated airline customer service scenario, it needs to understand the user's request, call tools in the system, and at the same time comply with the company's business rules. More critically, if the same task is executed repeatedly, can it keep getting it right? How far is a successful demonstration from reliably doing things on people's behalf?

Before personal assistants became popular, Noah was already studying two core questions about Agents: how to correct errors after they occur, and whether they can consistently finish things. This also explains why this young founder has attracted attention, secured a $1 billion Series C, and even raised three rounds within two months.

In terms of product design, Instinct chose an entry point we are very familiar with: text messaging. You can send it things to do just as you would contact a real person. For users, there is no need to build a new habit of opening an app; for the company, this also greatly reduces customer acquisition costs.

Howie Xu

Chief AI and Innovation Officer at Gen

I think future applications may very well be app-less. When you interact with it, it is just sending messages and conversing. What you need to tell it is the result, not what workflow to go through, clicking this and clicking that.

We have been spoiled by the previous generation or several generations of software. To book a restaurant or buy something, you have to click this and click that. Why should that be? You just tell it what you want, and let the model or software tell you what to do, rather than you doing it.

Turning an Agent into a contact that can help you get things done was not something Instinct thought of first.

In March this year, a product called Poke began exploring this path: connecting to email and calendars and putting the assistant into message conversations familiar to users. There is no need to open a standalone AI workbench; just send a message and let it handle the work.

But once the entry point becomes simple, a new problem appears: when faced with an assistant that can do anything, ordinary people may not know what they should ask it to do. In other words, an assistant having a certain capability and a user being willing to hand things over to it are two different matters.

Homin Luo

Founder of Mapier.ai

Instinct took off actually because there were many power users and pro-sumers on Twitter. These people first told you: "I used Instinct to do many things that Poke or GPT could not do before."

Many people found it very interesting, came to try it, and found that it could complete this use case. As a result, many very imaginative users went to try it, but not all users are very imaginative.

So specific use cases are very important for this kind of product. Early users sharing how they use it can be said to be helping the product do market education, but this also brings a difficult problem: can the traffic propped up by capital within two months of launch be sustained?

Howie Xu

Chief AI and Innovation Officer of Gen

In my view, Instinct is the new-generation Clubhouse. This judgment is not based on the product, but on the founder's or the company's playbook: it is high-profile and aggressive, just like Clubhouse back then.

There is nothing wrong with a high-profile, aggressive approach itself. Some companies succeed with it, while others stay relatively low-key. The problem is that when you go high-profile and aggressive, you need to be aware that you are doing so, be able to stretch and contract freely, and know when to extend and when to pull back. This is a big challenge for entrepreneurs.

After the novelty wore off, we rarely opened Clubhouse anymore, and Instinct faces the same problem. The founder has said he does not plan to make users pay for the product, which means making money through advertising. After the hype fades, is this business model viable? This is the biggest challenge facing startups.

Of course, being able to make so many people scramble to get an invitation is itself a kind of success. It is just that how far it will go after burning tokens funded by financing still has many variables.

Meta Muse's recent launch also brought a wave of attention to Instinct. The two products need to solve very similar problems, but they face different conditions: Instinct needs to make people aware of it first, while Meta already has a huge user network, namely existing social platforms such as Facebook, Instagram, and WhatsApp. This is the biggest difference between big companies and startups: distribution channels.

Howie Xu

Chief AI and Innovation Officer of Gen

Grok and Meta's Muse at least have distribution channels. Meta in particular has especially broad distribution channels. So I think in today's era, when a company has good distribution channels, it is hard for it to do too badly.

Meta Muse

Muse opened to the public on September 8 this year. As of early October, it had surpassed 6.6 million downloads and repeatedly reached the top of the App Store free app chart.

In addition to using Muse as a standalone app, people can also talk to Muse within WhatsApp and connect Facebook and Instagram accounts. This is Meta's natural advantage. It already participates in the process of users discovering content, developing interest, and connecting with others, and now it wants to further take on what needs to be done after these behaviors.

Take selling furniture as an example again. This is not a one-off task of writing a post, but a long-term delegated relationship. What I hand to Muse is a metric that requires long-term follow-up until the item is sold. This kind of task, on my Muse interface, is placed on a page called Goals.

According to Meta's introduction to Muse, what it wants to take on is exactly this type of task: understanding the user's goals, coordinating time and resources, continuing to push things forward after the user leaves, and coming back to contact the user when approval is needed or changes occur. It also remembers the user's preferences to provide reference for subsequent tasks.

To make this kind of delegation increasingly effortless, the assistant needs to remember what you have said before. Otherwise, every time you arrange a trip, you have to reintroduce your budget, accommodation preferences, and your family's schedules, and you still have to spend a lot of effort explaining the background. From a business perspective, this accumulation has the potential to create retention. Switching to another assistant means reintroducing yourself, going through the process of getting used to each other again, and rebuilding trust.

Meta's business vision for Muse is: for now, provide a relatively large free quota so users can try handing tasks to it; in the future, let Muse help users make money and save money, and Meta takes a cut from that.

Mark Zuckerberg

Chairman of Meta

We chose to do this because we are very confident that for people who use this product to run small businesses or engage in transactions and commercial activities, this product can help them make a lot of money.

In the long run, our business model will actually be very simple: just take a small commission from each transaction, making a big business very light. This commission may not even necessarily need to be paid by users themselves; it can also be borne by the merchants working with them.

To enable an Agent to establish this kind of long-term trusting relationship with users, what are the technical requirements behind it?

A sense of boundaries is one capability that the Muse team pays special attention to. For example, if you tell the Agent that you are pregnant, when it helps you book a restaurant, it can use that to look for places that offer suitable food and non-alcoholic drinks. But this does not mean it needs to tell the restaurant that you are pregnant as well. It needs to be able to use the information it knows to help you get things done, while judging which information does not need to be disclosed externally.

Why is Meta willing to invest so much in Muse? This brings us back to distribution, mentioned at the beginning of this chapter. Meta has already mastered content distribution in social media, and now it wants to further master task distribution. Specifically, it wants to become the first stop for users when they use software and purchase services.

Suppose you ask Muse to arrange a trip. In the past, you opened several platforms yourself, compared hotels, and then decided where to place the order. Now, you first tell the Agent the time, budget, and preferences; it searches, filters, and then hands the plan to you for confirmation. In this process, the assistant has the opportunity to influence which hotels enter the candidate list, and also has the opportunity to participate in the subsequent booking. If it does indeed bring orders to merchants, Meta then charges the merchants a fee—this is the business logic Zuckerberg referred to.

For a company like Meta that excels at advertising and precision targeting, an Agent can bring even more unique value.

Xiaoyin Qu

Founder and CEO of AgentSky

I used to work at Facebook. Facebook ads are essentially interest- and impulse-driven advertising: you see something while scrolling, act on impulse, and buy it right away. Google is more about having a strong need and going to search. So I think Muse's advantage for Facebook is that in many cases it can capture this strong intent very clearly. Before you even think of going to Google to search, Muse has already chatted with you and solved the matter—equivalent to bypassing Google's ads and solving it directly on Facebook.

The profit model Zuckerberg described is that I buy things on your behalf and then take a percentage commission. I think that even without buying things on your behalf, if Facebook understands my emails, contacts, and all my intentions, and then recommends feeds and ads to me, it will also become more precise and conversion rates will be higher.

Suppose the conversion rate of recommendations doubles, then revenue also doubles. For Facebook, this is a lot of money. By comparison, the compute itself is nothing.

From this perspective, it is easy to compare Instinct and Muse. Both hope users will form a habit: when there is something to do in life, first turn to this assistant. The difference is that Meta has the advantage of social apps, while Instinct needs to accumulate users, so it chose SMS, an entrance everyone is more familiar with.

However, these are still scenarios in daily life. In work scenarios, you may make another kind of request, and a very important aspect of that is division of labor. You ask AI to organize client materials, while also asking it to do market research, and also asking it to follow up on reimbursements. These tasks each have their own progress, files, and details that need to be remembered. Put them all into the same conversation, and once there are many things, you have to spend time organizing again.

Grok Bot, launched by SpaceX in the summer, is a multi-Agent collaboration product starting from the perspective of division of labor: organizing long-term delegations into a digital team.

Grok Bot

In the interview, Peng Zheng used autonomous driving to explain why humans become the bottleneck in Agent work.

Peng Zheng

Designer at SpaceXAI (Grok Bot)

Now, most people in this industry have already entered the stage of directing a group of Agents, or even a team of Agents, to do tasks, but everyone quickly realizes that we ourselves are the bottleneck. It is like autonomous driving: if you have to keep your hands on the steering wheel the whole time, it does not have much value for making use of a person's time. You still have to sit in the driver's seat, and your attention cannot be placed on other things.

The same is true in the digital world. We direct Agents to accomplish things, and at the execution level, they may already have replaced us in doing it ourselves. But if I still need to keep watching it and complete things through repeated back-and-forth, then it only puts our time into the role of manager and does not truly free up time.

So we hope to abstract the role and turn it into your digital colleague. Through this form, it may be possible to push human-Agent interaction toward greater automation and longer horizons.

People who have used Grok Bot will notice that its interface design is somewhat different from other Agent products. On the left side of the chat box are multiple Agents, or Bots, each with their own names, avatars, and division of labor, much like the way we communicate with a team at work.

These Bots can use computers in the cloud, enter work tools, pass information to each other, and collaborate in group chats. For example, one is responsible for client follow-up, one is responsible for marketing content, and another helps you coordinate tasks.

This design challenges the default form of our conversations with AI. In almost all AI products, we see three panels: on the left is chat history, in the middle is the current conversation, and beside it are tasks and results. When there is a new task, we open a new conversation, explain the requirements, and then repeatedly check, supplement, and confirm. But after these conversations end, we rarely look back, let alone organize the experience from them and leave it for the next task.

In this mode, although AI can already execute on our behalf, people still need to continuously push the work forward. Once there are many tasks, we change from people doing the work ourselves into managers constantly switching conversations and checking progress. Human attention, and the waking hours of each day, thus become the bottleneck for Agents to continue working.

So Grok Bot tries to change the organizational unit of the product from a single conversation to a long-existing Bot.

Peng Zheng

Designer at SpaceXAI (Grok Bot)

We turned the task panel on the left into a list of Agents centered on roles and identities. Through this interface change, we push users to think: when they receive a task, their first reaction should not be to open yet another new conversation, complete the task themselves, and monitor the results, but rather how to hand it off to their digital colleagues.

How do you automate tasks? How do you, in the process of completing tasks with it again and again, accumulate experience? What I want to design is probably not a one-off back-and-forth interaction, but a system that can automate as much as possible and omit the human factor from this loop as much as possible.

Multi-Agent collaboration also raises a very interesting design question. In human teams, we divide work because different people have different expertise. But in an Agent team, why still separate out a designer, an engineer, and a project manager? The key is that the things they need to remember are different.

Peng Zheng

Designer at SpaceXAI (Grok Bot)

This is more about how I limit context to reasonable boundaries, how I define where its distinctions actually lie—in practice, encapsulating the relevant context within the same Bot. All Skills and plugins are shared by every Bot, but all memory and personalization settings are separate; each Bot has its own memory and preferences.

The context here is the information an Agent references when executing a task, such as your requirements, previous conversations, and the work experience it has accumulated. For example, a design Bot needs to know what visual style you like; a Bot responsible for customer communication needs to remember how you interact with customers. If all this information is put together, it requires constantly judging: which requirements apply to the task at hand, and which are just special cases from another project. This problem of information interfering with each other is "context pollution."

So roles also become a way to organize memory. If you continuously hand one type of work to the same Bot, it can accumulate experience around that type of work and reduce the burden of re-explaining things each time.

This division of labor also does not need to be fixed. You can pull several related Bots into a group chat and let them collaborate around the same project; the same Bot can also participate in different projects at the same time.

But there is also a problem here: if we have a digital team, won't we then spend a lot of time managing it, so the human bottleneck remains? Therefore, another part of Grok Bot's design is to minimize what users need to learn and configure.

The team once tried to make the Agent's computer and operation process more visible. But in user research, they found that what many people want to confirm is: has the task been received? Is it still running now? Is it stuck? When do I need to come back?

So they use different animations of the avatar to express states such as the Bot working, waiting to be processed, or already completed. These seemingly small designs carry a very practical task: giving users enough information so they can leave for a while with peace of mind. The relationship Grok Bot wants to build is that you know it is still working, and you also know that when a judgment is needed, it will come back to you. The more detailed process can be expanded when you leave.

If this kind of cooperation can really be established, users' work habits will undergo a deeper change.

Peng Zheng

Designer at SpaceXAI (Grok Bot)

I think one thing that is happening now is that a layer of Agents is beginning to be added between people and the digital world, as if there is an extra layer of separation. For example, in my personal daily work, I often no longer personally use SaaS tools like Figma, Notion, or Slack, but instead direct my Agent to use them.

This also brings us back to the main thread of this issue: the entry point. Just as in life we use different software to book flights and find housing, at work we also have different SaaS software. All Agent products are subtly getting users to use these software through them.

So, from Instinct and Muse to Grok Bot, they are competing for a similar position, but their entry approaches differ. Instinct hopes to become the assistant in the message list that you can contact at any time; Meta hopes to leverage its existing social ecosystem to further take on the things users need to do; Grok Bot starts from the division of work, making different digital colleagues objects to which tasks can be continuously delegated. In the end, they all hope to establish the same habit: when there is a task, come to me first.

But when Agents really stand between users and software and merchants, new problems also arise. Can users safely let go? Are the original platforms willing to let them operate on users' behalf? And who ultimately bears the supervision, execution, and coordination costs that have not disappeared?

Hidden costs: privacy, security, and conflicts of interest

As more and more people use Agents, various "horror stories" have also appeared in media reports. For example, when asked to check flight terms, Instinct directly canceled the flight, causing a $300 loss; when asked to sell a keyboard, Muse, without permission, had the buyer come to pick it up, and the buyer, not seeing anyone, ultimately gave a negative review.

These cases cannot directly tell us how high the probability of an Agent making an error is, but they expose a very concrete problem: without our realizing it, how much authorization have we actually given the Agent? The first hidden cost of an Agent is the privacy and security boundary.

Howie Xu

Chief AI and Innovation Officer at Gen

I think the most important thing is probably still security and reliability, because this is what users easily overlook. When something new first comes out, everyone cares about how many appealing features it has and what it can do: "I just say one sentence, and it books the flight for me." But for something new to truly enter millions of households over the long term, the requirements for security and reliability are certainly still relatively high.

There are two different issues here: one is whether the assistant will act beyond your authorization; the other is who will see the data you give it, how it will be stored, and how it will be used.

Zuckerberg revealed that he and Nat Friedman personally recruited Moxie Marlinspike, the founder of the encrypted messaging app Signal, to have him participate in Muse's confidential virtual machine project. This is to put the information users give to Muse into a protected cloud runtime environment, with the goal that even Meta itself cannot view the contents inside.

Meta and Zuckerberg's recent actions are also quite intriguing. Due to the incidents at Hugging Face and OpenAI, public attention to AI safety has recently been high. At this point, Zuckerberg posted a tweet saying that to ensure Muse's safety, Meta would delay the release by several months compared with the original plan.

In addition, Meta also added a monitoring Agent called Sentinel, which checks information going in and out and triggers user review when needed.

But even if the question of who can see the data is solved, the question of what the assistant will do with the data is not fully solved. It may still, in the course of a normal task, send a message that should not be sent, or perform a task you have not approved.

Homin Luo

Founder of Mapier.ai

We can constrain ourselves at the infrastructure layer, but we have no way to constrain our customers: what they will do once they use this user data.

The second issue is whether the original platforms are willing to let Agents in.

Not long after Muse became popular, Amazon blocked Muse, prohibiting it from shopping on users' behalf. Amazon said that Muse had not publicly identified itself as an Agent and raised security concerns about account access and credential handling. It believes that third-party services that purchase things for users from other merchants should operate publicly and respect the merchant's decision on whether to participate.

The disagreement here is: if a user agrees to let an assistant operate on their behalf, is that enough for another platform to accept this assistant? From the user's perspective, I have an account and I am willing to spend money; I have just changed the way I operate.

But from the platform's perspective, the one entering the account, reading information, and executing transactions also includes a system provided by another company. It wants to know who this system is, what it will do, and how problems will be handled if they arise.

Beyond security and access rules, there is also a layer of conflict over commercial interests. Meta hopes to move further from content distribution to task distribution, but these tasks originally take place on other platforms. If you open a shopping website yourself, the platform can influence your choices through search, recommendations, and advertising. If Muse first compares products and then completes the purchase, part of the selection process may shift to Muse.

The original platform may still receive orders, but it cannot directly reach users and influence purchase decisions as before, and advertising revenue may also decline. This is the commercial pressure visible in this conflict: while the Agent brings transactions to the platform, it may also take away the customer relationship.

Privacy and security concerns and platform conflicts make the idea of one Agent taking over all tasks more complicated, especially at the enterprise level.

After the entry point, where will the future of Agents go?

At this point, there is one last question: can these Agents ultimately become a profitable business?

When we were writing this article, there had already been many reports on user growth, financing, and valuations, but we had not yet seen clear revenue and profit disclosures for Muse, Instinct, and Grok Bot respectively.

In analyzing Instinct and Muse, the two Agents mainly targeting the consumer side, we have already mentioned that their profit prospects depend entirely on existing distribution channels.

For a company like Meta that already has mature businesses, it can choose to invest first, let users form habits, and then gradually look for sources of revenue. Independent startups like Instinct can also subsidize usage through financing, but whether the next round of financing can continue will ultimately still have to answer questions about growth, retention, and the business model.

Xiaoyin Qu

Founder and CEO of AgentSky

I think consumer Agents ultimately come down to who has more money and can burn it. Even if Instinct's experience is better, if it cannot be 100% free, I think it has no future.

In the Agent business, there is also a particular contradiction: the more users like it and the more willing they are to hand things over to it, the more times the model must be called, and the company's service costs may increase accordingly.

Xiaoyin Qu

Founder and CEO of AgentSky

So I think that, in the end, you'll find that this spending grows exponentially. I don't think this is a game a startup can play.

This also explains why many Agent products enter through office scenarios.

Peng Zheng

SpaceXAI designer (Grok Bot)

Office scenarios have clearer deliverables, such as following up on a PR or making a weekly briefing and coordinating schedules. Enterprise users have a stronger willingness to pay and can better prove this kind of demand: I delegate a task and get the result directly.

If an assistant can reliably complete weekly reports or continuously handle customer follow-ups, it becomes easier for enterprises to measure: how many work hours does it save, and is it worth paying for? But a stronger willingness to pay does not mean the money is easier to earn. Data protection, permission management, and system integration may all increase costs for enterprises.

Either way, this battle over Agent products has only just begun.

Instinct, Muse, and Grok Bot—some want to become personal assistants, while others want to form a digital team. Their product forms and entry points differ, but the core logic is the same: seize the entry point through which people execute tasks. That entry point used to be software; today it is the Agent, which wants to become the bridge between us and the digital world.

Of course, today's form is not necessarily tomorrow's answer. One possible change is the birth of specialized Agents. What we see now are still general-purpose Agents; in the future, there may be Agents focused on niche areas, serving the different needs of enterprises and even consumers. Today's diversity of models may be tomorrow's diversity of Agents.

Xiaoyin Qu

Founder and CEO of AgentSky

Each employee may use the same production line, possibly with multiple Agents. How these Agents cooperate and how the production line is built—I think this has only just begun. Around these questions—what kind of Agent does what under what circumstances, when we care about price, when we care about quality, and when we care about iteration speed—these things will gradually emerge. We are still at a very early stage.

But the high valuations of companies like Instinct already reflect capital's bet on one possibility: once we get used to first telling our needs to an assistant, which then calls software and finds services, the assistant has a chance to become the new entry point.

Whoever can win this long-term trust will have the chance to change the way we use software and buy services. This is also why we will continue to follow this competition.