Summary of Key Points
This article criticizes the current overemphasis on the "Agent + Skill" combination model in the AI field, arguing that this approach merely shifts the complexity of development to the users (transforming the task from "learning to use software" to "learning to assemble software"). It emphasizes that what users truly want is not the tools themselves, but the direct outcome of getting tasks completed. The article proposes that "demand intermediaries" will be the key entry points for traffic in the AI era, as they can transform users' natural language queries into actionable plans and coordinate various resources to complete tasks. It also highlights that giants with control over system permissions, models, and user data will dominate this field, creating barriers through these intermediaries (reducing users to mere access to resource indexes rather than core capabilities). Ultimately, it asserts that the true goal of AI is to be responsible for the outcomes, not the tools or mechanisms themselves.
I. The Agent Model: Why is it like "teaching passengers to fly a plane"?
Many current AI products require users to create their own Agents, configure Skills, and integrate knowledge bases. This sounds sophisticated, but in essence, it shifts the work that should be done by development teams onto the users. For example, if you want to filter resumes, you used to just ask for "help in finding eligible candidates," but now you have to choose an Agent, select appropriate Skills, and adjust parameters—similar to buying a plane ticket and then being expected to learn how to fly it yourself before the flight can depart. Users want to reach their destination (a filtered list of resumes); who cares about how many blades the engine has? This model makes users act like product managers, adding unnecessary complexity and going against the AI mission of simplifying life.
II. "Demand Intermediaries": The "Super Secretaries" of the AI Era
The author refers to "demand intermediaries" as entities that understand user needs and handle everything on their behalf. For instance, if you ask for help in identifying risks in a contract, they would:
1. Convert your natural language request into clear tasks (such as "identifying breach of contract clauses or loopholes in the disclaimer");
2. Search for suitable AI tools in an "Agent store" (without you having to do it yourself);
3. Combine these tools to complete the task;
4. Provide a clear result (e.g., "There is a risk in clause 3 of the contract"). These intermediaries are like browsers in the internet era or app stores in the mobile era—connecting users with content. They could be hidden in voice assistants, input methods, or even small icons on desktops, but their core function is to allow users to simply state their needs without worrying about the process.
III. Why Can Giants Dominate "Demand Intermediaries"?
While independent AI products may start as prototypes of demand intermediaries, giants are likely to become the winners. The reasons are straightforward:
- System Permissions: Built-in assistants on smartphones (like Apple Siri or Huawei Xiaoyi) have access to more phone functions (cameras, contacts), which third-party assistants don't. AI giants (such as OpenAI or ByteDance) control large models and operating systems, enabling better integration of resources.
- Data Advantages: Giants possess long-term user behavior data, allowing them to understand needs more accurately; independent products struggle to access this comprehensive information.
- Resource Integration: Giants can seamlessly integrate their own models, tools, and knowledge bases, providing a smoother experience than independent products. Past examples like 91 Assistant and PeaPod were eventually acquired by smartphone manufacturers; the same will likely happen with AI demand intermediaries.
IV. How Platforms Use "Demand Intermediaries" to Build Barriers
Platforms (such as large model companies) use demand intermediaries to lock users into their ecosystems:
- Resource Indexing: Skills provided to users are not complete capabilities but references to resources (e.g., "using the platform's web search tool v3.2, fact-checking policy v2.0"). It’s like receiving a list of ingredients, but all the materials are stored on the platform; if you switch platforms, the list becomes useless.
- Unified Maintenance: Platforms manage common rules (e.g., fact-checking standards) centrally, updating them for all users at once and reducing redundancy. Personal data is kept separate from platform capabilities, ensuring security while giving the platform control. This creates a dependency, making it difficult for users to leave the platform.
V. The Final Goal of AI: Whoever Can Be Responsible for the Outcomes Wins
The article concludes that the ultimate success in AI will not depend on having the most impressive Agents or Skills, but on being able to take responsibility for the outcomes. For example, if you use AI to filter resumes, it must provide accurate results and be accountable for any missed candidates; if you use it to identify contract risks, it must ensure no critical issues are overlooked. In the software era, competition focused on features; in the SaaS era, on smooth workflows; in the Agent era, on execution capabilities. The winner will be the one who can make users trust them with their tasks.
In summary, this article warns us not to be misled by new terms like Agents and Skills. What users really want is for tasks to be completed efficiently. The future of AI lies in having demand intermediaries handle everything, with giants taking responsibility for the outcomes.