The race for personal agents begins before users even realize it.

Before booking a restaurant, you may have gone through various thoughts—wanting to meet a friend, celebrate, or simply change the environment; the same phrase “I want to go out tonight” can ultimately lead to a meal, a performance, or even a route with no predetermined destination.

In other words, consumption begins before specific details like store name, category, or price appear. At that stage it has no name yet, resembling a vague feeling or an unformed desire. Search, content recommendations, and friends’ suggestions all influence this process.

MachinePulse‘sKarpo, an AI product focused on urban life that makes choices at this stage.

Starting from a mood: choose the right state, express the feeling you want, and Karpo can suggest specific destinations. Image source: Karpo

It and Resy, Google Maps are not on the same starting point. The latter deals with tasks after you already know what you’re looking for—searching locations, checking ratings, planning routes, and so on; whereas Karpo focuses on helping users who haven’t yet decided on a destination, or even on what type of activity they want to do, move from a vague idea to a concrete plan.

Behind this choice is a team with consumer‑grade AI product experience. Founder Leah, was an early member of Doubao, participating in the entire process of Doubao growing from 0 to 1 hundred million DAU. Co‑founder Allen previously worked as an engineer at Doubao and TikTok, and has multiple entrepreneurial experiences; another co‑founder Alex was previously the head of the Jump Agent product team.

With this experience, the team chose to explore the overseas market“Lifestyle×AI”. In their view, serving an individual requires being close to specific scenarios, understanding personal taste, and also dealing with real‑world supply. This provides a concrete entry point for consumer‑grade Personal Agent.

So, when AI can intervene in decisions before users form explicit needs, Personal Agent— what changes will this competition bring about?

From searching for answers to participating in requirement formation

First, we should be clear,Personal Agent brings a key change: besides being able to handle more tasks, it also turns the“whether the requirement is clear” problem that was previously assumed to be solved by the user themselves, into an issue that the product must handle.

This actually reflects a division between two capabilities. Execution ability addresses how to get something done: once the goal is set, the rest involves breaking down steps, invoking tools, and advancing the process; whereas understanding intent tackles a more upstream problem, helping the user gradually clarify exactly what they want. The latter is especially common in everyday decision‑making scenarios.

Life experiences often blend emotions, aesthetics, relationships, and spontaneous impulses. People may already anticipate a certain feeling but still don’t know which activity or place can satisfy it. The need isn’t absent; it just hasn’t been organized into a directly actionable requirement.

This understanding also shows up in information that isn’t repeated. For example, when askingKarpo asks:“I want to run in a quiet place. Where is suitable?”There is no mention of dogs, nor of cars.

ButKarpo‘s answer begins with:“you andKodagoing to the riverside is absolutely fine.”Kodais the family dog,Karporemembers. It recommended two quiet, dog‑friendly, water‑accessible spots, then added:“Sportagecan be driven there without issue,but the car’s oil needs to be changed, which doesn’t affect today’s round‑trip.”

You don’t have to explain your life to it every time; that’s the most genuine feeling after receiving a reply.

Of course, this level of interaction only occurs after personal understanding has accumulated to a certain point: the basis is not just the current message but the context built from previous dialogue. Intent capture doesn’t happen only in a single Q&A; it requires an ongoing relationship as a foundation.

And to do this well, merely remembering what you like isn’t enough.

A user who just arrived in New York tells Karpo about their interests; Karpo continuously recommends places frequented by similar people, helping them find regular spots and friends in the new city.

Someone who has long liked art does not necessarily want to go to a museum every weekend. Past preferences provide a basis for understanding, while current mood, peer relationships, and schedule determine how the recommendation should be adjusted. Even for the same exhibition, when going alone you might be willing to spend half a day studying the works, whereas with friends you may care more about whether both parties are interested.

Feedback also needs to be understood in its specific context. A user may skip an exhibition because the content doesn’t match their interests, or simply because it’s too far away. These two reasons provide different bases for the next recommendation. Understanding an individual must accommodate such changes, distinguishing stable preferences from one‑off decision factors.

“Proactive experience” is also built on this understanding. Sending a proactive message is merely an interaction format; the value of the suggestion depends more on the context, timing, and whether it helps the user move to the next step. The product needs to, based on conversation, preferences, and specific scenarios, present relevant options at the appropriate moment, allowing the user to confirm, adjust, and gradually form a plan.

The starting point of intent competition is to reduce the cost for users to express needs and make choices. Whether that choice can truly become an experience also depends on another factor:AIhow to connect an understanding of a person with the real supply that exists in the city. 

After understanding a person, you also need to reorganize the world in front of them

In fact, making a vague intention clear does not equal completing the service. What’s nearby, when it opens, whether it’s suitable for peers, and how transportation and activities connect all affect whether a suggestion can become a real arrangement. No matter how accurate the subjective feeling, it ultimately must stand up to the test of time, space, and supply conditions.

This means,AI needs to understand not only what is worth recommending, but also the relationships between different options. How long a place is suitable to stay, how far it is from the next destination, and whether the overall pacing of the itinerary meets user expectations all affect the final experience.

After confirming the requirements, Karpo compares prices, verifies refund and change policies, and facilitates ticket purchase, handing it over to the user for confirmation before placing the order.

A particular type of demand better tests this capability: when the time window is compressed,AIcan organize city supply into a rhythmic route. For example, sending a message during a flight layover:“I1:10land in Barcelona,9:20the next flight—can I do anything in between?”

Karpo turns a 8‑hour constraint into a closed route: first go to El Born, follow Santa Maria del Mar to Port Vell, then continue to Barceloneta beach, linking the old town with the sea view to avoid backtracking; then at Can Solé reserve a table in advance, “otherwise the transfer time will be wasted waiting for a seat”; in the afternoon head to Montjuïc to watch the sunset, 6:45 depart, head straight back to El Prat airport.

At this point you might wonder:“Like, can I see the sea, eat seafood paella, and still catch the sunset?”Karpo‘s answer is:“and you can still catch your flight.”

The focus is not onKarpo recommended which places, but rather on the order and timing each place appears, all arranged around the transfer window: the route reduces backtracking, restaurant reservations reduce waiting, and the return leg serves to catch the next flight.

This is exactly “supply organization” and “information retrieval”‘s difference: retrieval is finding a suitable option, whereas supply organization combines multiple originally scattered merchants and locations according to an individual’s needs and rhythm, allowing them to jointly satisfy a complete life arrangement.

Once you figure this out, you’ll understandKarpoWhy choose urban life as the angle.

Eating, drinking, playing, and entertainment may seem like light scenarios, but they also involve complex issues such as personal taste, local information, and activity planning. Users often don’t lack options; they just don’t know which choice is best for them at the moment. The richer the content a city offers, the more effort is required to filter and combine it. The more venues and activities there are, the process of finding a suitable arrangement isn’t necessarily simpler.

The supply side is the same. Merchants have offerings, but they may not be discovered by the people who exactly need them; a restaurant has its own atmosphere, an exhibition has a specific expression, and these characteristics are not always fully captured by tags such as ratings, categories, or popularity.

AI opportunities exist between these two ends.

Whoever embraces the intent has the opportunity to influence the direction of consumption

When AI participates in demand formation, and further organizes selection and action, it has the opportunity to change the relationship between consumption entry points and distribution; this is the business significance of intent competition that deserves industry attention.

As mentioned earlier, the destination of a purchase is often influenced before the order is placed,Personal AgentWhat it explores is to further connect the understanding of intent with subsequent arrangements through continuous communication centered around the individual.

If a person could start a life arrangement without first defining a category, destination, or merchant, then the need might cross existing classifications during AI communication, the demand can transcend traditional categories. From wanting to unwind, to deciding to visit an exhibition, and then arranging a nearby restaurant, several options that originally required separate searches and comparisons can unfold around a single idea step by step. The supply‑side logic also changes accordingly, and services have the opportunity to shift from offering a single product to fulfilling an entire chain of needs.

In the era of search and reviews, whether a business or an event can be discovered largely depends on whether it matches the keywords users enter, has a high enough rating, and is positioned prominently. AI organizes around a specific person, it relies even more on how well it aligns with a particular user and a specific context.

This will bring a significant shift: experiences that were hard to describe with keywords and therefore difficult to discover may now enter someone’s consideration for the first time because they precisely match an individual’s interests. The visibility of supply will no longer be determined solely by the search box.

Seeing the weight of this position, capital has already started to bet early.

Currently Karpo has received support from Creekstone Ventures, Hillhouse Capital, BlueLake Capital and other firms, raising a total of tens of millions of dollars.

Put it into the wholePersonal Agent‘s track, the meaning of this action becomes clearer: whenMuse,Instinctthese kinds of products are fiercely competing at the task execution level, there is also a group of capital and entrepreneurs willing to invest in the upstream“intent competition”pricing.

What they’re betting on is whether the consumer entry point will shift from the moment you already know what you want to the moment you haven’t even thought of wanting anything yet. This leads to a deeper competitive logic beyond whose recommendation is more accurate.

A single isolated recommendation is easy to replace; users can switch tools and ask again at any time. What is truly irreplaceable is an ongoing relationship.AIUnderstanding a person’s preferences over the long term and continuously handling subsequent actions is a self‑reinforcing process. User feedback makes the suggestions increasingly accurate; more accurate suggestions increase their willingness to delegate tasks, and more delegation brings richer understanding and data.

In other words, the real competitive metric here is how many tasks and how early users are willing to entrust to this AI, rather than the win‑loss of a single‑turn Q&A.Karpo‘s exploration around personal taste and urban supply is precisely the accumulation of such a relationship, and it also presents a possible “Lifestyle×AI” as a possibility.

During the New York Marathon, Karpo partnered with local New York running groups, supporting on-site events and integrating into the city’s running community.

When intentions can be captured earlier and more smoothly, the form of consumption will change accordingly. A large number of demands that previously never materialized because they were hard to articulate, people were too lazy to seek, or feared the hassle are now being released.KarpoThis is only the first step; an era where more ideas can be seen and realized is unfolding.

Depth, insight, foresight, capturing business opportunities for the top 1%

Curated official accounts of 36Kr

👇  Sincerely recommend you to follow 👇