Let me ask you a question, have you heard of or used more than half of the Agent products mentioned in the picture below?
More than a third?
Over the past 25 years, in less than half a year, I have witnessed and personally tested several excellent Agent products, such as: Manus, Flowith Neo, Genspark, Lovart, Lovable, II Agent, Skywork Super Agents, MiniMax Agent...
There are other things that I don’t use much or have never used but know: There are more that I haven’t heard of: Based on past experience, I speculate that more than 90% of the Agent products currently on the market will be eliminated.
Especially now it is still in the early stages of the Agent ecosystem.
I heard from a guest on a podcast that a certain factory is developing four AI Agent products similar to Manus at the same time.
Let those poor children go📷 In the past half month, I have gradually become a little PTSD from various Agent products, and gradually returned to the interactive window of Gemini and Calude.
However, Agent products still pop up one after another.
Whether you use them or not, they still pop out stubbornly, and I just want to be quiet.
In fact, I have been thinking about a question in the past two weeks: "How to design an AI Agent product with continuous payment value in an Agent market that is constantly evolving in basic capabilities and has serious functional homogeneity?
"Whether it is a vertical Agent or a general Agent product, the ability to pay continuously for a long time is the most direct criterion for testing whether the product is valuable.
The following thoughts are highly subjective, and I do not seek approval.
They are just to inspire others.
I hope to provide you with some ideas.
A pure little world: don’t be greedy for too much, be distinctive, and be pure.
The common crux of current Agent products is “greedy for the big and seeking the complete”.
It’s tempting to cram all the functions into one product.
The result is often comprehensive but mediocre.
It's like a cafeteria with plenty of dishes but none of them are memorable.
A truly valuable product must be a "pure little world" that focuses on solving the specific pain points of a specific user group and makes it the ultimate.
I have a bold idea, why not just abandon those "universal features" that seem to appeal to a wide range of users?
Instead, all R&D resources and efforts will be invested in polishing core competitiveness.
For example, the value of an AI Agent focused on legal consultation does not lie in its ability to "chat" or "write poetry", but in its precise understanding of legal provisions, unique insights in case analysis, and impeccable writing of legal documents.
This purity is precisely the key to building user trust and forming payment barriers.
Imagine that when users face a complex and time-sensitive legal issue, will they choose a "universal" Agent with complex functions and general responses, or a "pure" Agent that can provide highly professional and accurate legal advice?
This purity makes the product no longer a "dispensable" tool, but an "indispensable" expert.
Product content and interactions that are accurate, unique, and suitable for user characteristics.
The value of AI Agent lies in its ability to provide highly customized content and interactive experiences based on user characteristics, habits, and preferences.
Think about what we mentioned before about thousands of people and thousands of faces.
These are the big Internet companies that want to build a lot of machines, data, and algorithms.
Now that there is LLM between products and users, small but beautiful products also have the opportunity to provide thousands of services to thousands of people.
This is not just a simple "recommendation" function, but also requires in-depth "understanding" and "prediction" of user intentions.
For example, an AI Agent in the education field should not just be a builder of question banks and courses, but should be a "private tutor" who can understand students' learning styles, knowledge blind spots, and emotional ups and downs.
The learning content it provides is not a standardized course.
It requires a personalized learning path that is dynamically adjusted according to the student's current cognitive level, points of interest, and even emotional state.
Interaction methods should also be diverse, ranging from serious knowledge questions and answers to relaxed conversational tutoring.
They can even sense students' emotional changes and provide encouragement or guidance in a timely manner.
The connotation of "suiting user characteristics" also includes the exploration of users' "hidden needs".
Many times, users cannot clearly express all their needs.
The intelligence of AI Agent lies in its ability to predict and meet these potential needs through the analysis of massive data and insights into user behavior patterns.
A project management agent can not only create tasks according to user instructions, but also proactively identify dependencies between tasks, warn of potential risks, and even proactively provide relaxation techniques or psychological counseling when users feel stressed.
This is truly humane and the key to its ability to establish deep connections with users and form ongoing paying relationships.
Can quickly integrate the latest capabilities of large models into the strongest functions of the product.
For AI Agent products, whether they can quickly integrate the latest capabilities of large models and deeply integrate their strongest functions with them directly determines their vitality in the market.
This requires that our product architecture must be highly flexible and scalable.
Rather than simply "grafting" new technologies, it can deeply understand the underlying logic of new technologies and "integrate them at the genetic level" with the core functions of the product.
When new multi-modal large models emerge, a text creation agent should not only support text input, but also be able to understand information from multiple modalities such as images and videos, and integrate it into the creation process to generate more expressive and infectious content.
This "rapid integration" is not just a simple API call, but also requires optimization and adaptation at the model level, and even redesigning part of the interaction logic to fully unleash the potential of the new technology.
At the same time, the continuous optimization and improvement of the product's strongest functions is also an important part of maintaining competitiveness.
This involves a refined data flywheel effect, continuous user feedback loops, and forward-looking investment in technology research and development.
Only by combining "fast" with "precision" can we seize the opportunity in the fierce technological competition and capture sustained value.
It may sound difficult, but do you have many options for a small business?
Rely on creativity, design concept?
AI Coding can be copied at any rate by relying on technological innovation?
A large model manufacturer with top-level voice can reproduce it within two weeks, and then take advantage of its traffic and price advantages to eat it instantly.
At present, it seems that AI products at the application layer can only quickly form die-hard fans of the product through constant trial and error combinations before the elephant turns around.
It's hard work and staying sharp and smart at the same time.
It is best to let users develop usage habits within three months.
In this case, the migration cost for users will still be quite high.
Also, try to let users retain personalized and precious personal data in the product.
At least I am not good at migrating this personal data because of the cheaper price.
I'd even pay extra for storage for it.
Agent capabilities do not necessarily have everything, and there is no need to pursue multiple agents.
Last year, I still held the view that multiple agents must be better than the single-agent model, and the more agents, the better.
Each Agent performs its own duties, collaborates with each other, and has a flexible architecture.
This design is awesome.
I was also obsessed with multi-Agent architecture.
I recently discovered that something was not quite right, and I also tried to use my own hands to write Agents, single-Agent frameworks, and multi-Agent frameworks to write similar functions.
Theoretical elegance is actually not tenable at all.
In the end, we have to go back to the design of product interaction, core function prompt tuning, continuous optimization of data processing links, how to maximize the reasoning ability of connecting large model APIs, multi-modal capabilities and user expectations, and the balance between the existing data in the system.
This has to be done bit by bit and debugged over and over again.
Elegant design becomes a heap.
It makes no sense.
The core value of an Agent product lies in its ability to complete specific tasks efficiently and accurately and solve user pain points.
If a single Agent can complete this task perfectly, then there is no need to introduce multiple Agents.
An Agent focused on schedule management, its core functions are schedule creation, reminders and conflict resolution.
If an "emotion-aware agent" is forcibly added to cooperate with the schedule agent in order to pursue "multi-agent", it may lead to increased system complexity, slower response speed, and even unnecessary interference.
The real wisdom lies in "simplification" and investing limited resources into the agent capabilities that can create the most value.
This requires us to have in-depth insights into user needs, a thorough understanding of business processes, and the ability to accurately evaluate the contribution of different Agent capabilities to product value.
If the introduction of a certain Agent capability can significantly improve the efficiency and accuracy of the product in solving problems, then it should be fully invested; on the contrary, if it is just to "look advanced", then it should be decisively given up.
This is a key decision that requires strategic determination and business judgment.
This decision-making ability can only be accumulated through practice.
This is the real barrier.
