The next wave to disrupt the internet: The Agentic Web is here.

You won't "surf the Internet" anymore. Instead, you'll state a goal, and a group of AIs will automatically accomplish it.

—— A vision of future Internet usage scenarios

Over the past three decades, the Internet has undergone a profound evolution from static web pages to intelligent recommendation systems. Today, we stand at another major turning point in the development of the Internet.

This transformation stems from a brand - new paradigm concept —— Agentic Web, a goal - oriented Internet system composed of AI agents. In this new framework, users no longer manually browse web pages or click buttons. Instead, they issue a goal to the agents through natural language, and the AI will autonomously plan, search, invoke services, coordinate with other agents, and ultimately complete complex tasks.

This is not a fantasy. It is a Web reconstruction plan jointly proposed by researchers from institutions such as UC Berkeley, UCL, Shanghai Jiao Tong University, and Shanghai Chuangzhi College, and systematically discussed in a research paper.

Paper Title: Agentic Web: Weaving the Next Web with AI Agents

Authors: Yingxuan Yang, Mulei Ma, Yuxuan Huang, Huacan Chai, Chenyu Gong, Haoran Geng, Yuanjian Zhou, Ying Wen, Meng Fang, Muhao Chen, Shangding Gu, Ming Jin, Costas Spanos, Yang Yang, Pieter Abbeel, Dawn Song, Weinan Zhang, Jun Wang

Affiliations: Shanghai Jiao Tong University, University of California, Berkeley, University College London, Shanghai Chuangzhi College, etc.

Link: https://arxiv.org/abs/2507.21206

Github: https://github.com/SafeRL - Lab/agentic - web

This is a comprehensive "rewrite proposal" for the underlying logic of the Internet: Humans are no longer the sole users of the network. Agents will become the main operators of the Web. Tasks are initiated by humans but executed by AI. In this new architecture, web pages, services, and platforms are no longer interactive interfaces for humans but collaborative interfaces for agents.

This article will conduct an in - depth analysis of this paradigm revolution of the "agent - driven Internet" from five aspects: technical architecture, theoretical model, system protocol, typical applications, and challenges.

I. Three Paradigm Shifts: The Web is Moving Towards "Automation"

The evolution of the Internet is a technological history of the "human - information" relationship. Over the past three decades, the Web has mainly experienced three paradigm shifts:

PC Web: The "Directory Network" Driven by Keywords

In the PC Web era, web pages mainly consisted of static content. Information was centrally generated by institutions and formed a "digital yellow pages" through manual classification and hyperlinks. Users had to actively initiate searches, click and browse. Task execution was linear, clear but not very efficient.

The business model was mainly based on keyword search advertising. Representative systems such as Google AdWords relied on click - through rate (CTR) and cost - per - click (CPC) to measure effectiveness, forming a search marketing ecosystem based on "human intent".

Mobile Web: The "Content Explosion" Driven by Recommendations

With the surge of social platforms, short - videos, and e - commerce UGC, the amount of information has increased exponentially. Traditional search engines have difficulty coping with the huge content distribution pressure. Instead, the information distribution paradigm is dominated by recommendation systems.

Users have gradually changed from "searchers" to "consumers". Algorithms dynamically recommend content based on behavioral data, and platforms have changed from content aggregators to algorithmic intermediaries. The business model has shifted to precise recommendation and in - feed advertising, emphasizing time spent, conversion rate, and effective cost per mille (eCPM).

Agentic Web: The "Action Network" Driven by Agents

Today, we are entering the third wave of change: AI agents are taking the lead, and the Web is shifting from "human - readable content" to "agent - executed tasks". Information is no longer statically stored in web pages but is embedded in LLM parameters, to be invoked, combined, and re - processed by agents.

The role of the Web is no longer an information warehouse but an ecosystem full of "actionable resources" for agents to discover, coordinate, and invoke. Tasks no longer rely on users' step - by - step operations but are completed by AI agents throughout the entire process, from information discovery to service invocation and result feedback.

This trend indicates that: the future Web will be built, operated, and used by AI agents. We need to re - understand what a "web page" is, what "traffic" is, and even what a "user" is.

The Internet is no longer just a space for humans. It is gradually becoming an ecosystem where agents participate, collaborate, and create value together.

II. What is Agentic Web?

The definition in the paper states:

Agentic Web is a distributed and interactive Internet ecosystem in which autonomous software agents driven by large language models (LLMs) can continuously plan, coordinate, and execute goal - oriented tasks. In this paradigm, network resources and services are not only available to humans but also accessible to agents, making agent - to - agent interactions the norm.

In short, it is a network form where AI "surfs the Internet" and executes tasks, while humans only "issue instructions".

The Core of Agentic Web is "Delegation + Execution"

In Agentic Web, users no longer need to manually search, click, copy, or paste content. Instead, they can delegate tasks through conversations with agents. For example, users only need to say:

"Help me plan a weekend trip to Tokyo with a budget of 3,000 yuan and avoid typhoons."

After that, all the remaining work will be automatically completed by the agent —— from querying the weather, searching for flights, comparing prices, to booking hotels and integrating schedules. The entire process is fully automated. Moreover, these agents can collaborate and negotiate with other agents (such as airline APIs, hotel APIs, and travel data agents) to achieve task goals. This is not just a single - round Q&A like ChatGPT but is completed through multiple steps and multi - agent collaboration, representing that AI is truly involved in the operational level of the Web.

The identity of an agent in the system is "dual":

Agent - as - User (As a User)

Just as humans access web pages, agents can simulate clicks, fill out forms, read interfaces, and perform tasks such as market analysis, data scraping, and automatic trading.

Agent - as - Interface (As an Interface)

Agents can also act as "super assistants", receiving natural language instructions from users, automatically parsing, invoking multiple services, integrating results, and executing multi - step processes.

A complete agent often has both roles: it can represent humans to interact with the system and also serve as an interface for the system to humans, truly achieving a closed - loop of "intention - execution".

III. Understanding the "Three Core Dimensions" of Agentic Web

The paper comprehensively understands the structure of Agentic Web from three core dimensions:

Intelligence Dimension AI agents need to have real "cognitive abilities", including:

Context understanding: being able to read web pages, structured data, and natural language

Long - term planning: being able to break down complex tasks and generate execution plans

Adaptive learning: continuously optimizing strategies through experience

Multi - modal integration: simultaneously processing text, images, APIs, data tables, etc.

These abilities mean that agents are not passive "response tools" but "digital actors" with continuous learning and autonomous strategies.

Interaction Dimension Agentic Web breaks the operation paradigm of "humans clicking on web pages" and shifts to semantic - based intelligent interaction:

Using MCP (Model Context Protocol) and A2A (Agent - to - Agent) protocols to achieve agent discovery, ability description, and state sharing

Supporting multi - step task context maintenance (such as shopping processes and medical consultation processes)

Realizing Agent - to - Agent collaboration and task decomposition

Agents do not "invoke" each other but negotiate and collaborate to execute. For example, a travel agent actively requests data from a weather agent and then links up with map and ticketing tools to complete a task.

Economic Dimension

The most groundbreaking idea in Agentic Web is: Agent Attention Economy

The traditional advertising model pursues "human clicks"; in Agentic Web, the objects that resource providers compete for become "the invocation of AI agents".

This means that in the future, there will be:

Recommendation systems for agents;

Advertisements targeted at agents;

Bidding in the service market based on "agent invocation rate";

The invocation frequency, completion rate, and efficiency of agents will become new "traffic indicators", and the focus of business competition will shift from competing for user attention to competing for agent "attention".

IV. Application Scenarios: From Search Substitution to Intelligent Transaction Systems

To better understand its practical value, we can break down the core capabilities of Agentic Web into three categories: transactional, informational, and communicational. They together constitute three basic ways for agents to participate in the digital world.

Transactional: From "Click to Order" to "Fully Automated Task Completion"

In the traditional Web, users need to browse page by page, search for information, and perform step - by - step operations to complete a task, such as booking a hotel, buying a plane ticket, or applying for a visa. In Agentic Web, you only need to tell the agent:

"Help me book a round - trip ticket from Shanghai to Tokyo next Wednesday in economy class, avoiding typhoons."

The rest —— querying airlines, comparing prices, confirming the time, filling out information, and confirming payment —— will all be completed autonomously by the agent. It not only invokes airline APIs but also weighs according to your past preferences (such as credit card points and eco - friendly routes), and can even automatically re - book in case of changes.

This intelligent transaction processing ability is being further extended to the device layer by "Mobile Agents" and "App Agents". For example, agents can synchronize schedules on your phone, modify meeting arrangements, and even integrate multiple applications to automatically execute cross - platform tasks.

Informational: From "Search Engines" to "Continuous Knowledge Discovery"

Today's information retrieval relies on search engines and social recommendations. However, in the context of data overload, we are exposed to a flood of information.

The "informational agents" supported by Agentic Web are more like long - term research assistants. Take the "Deepresearch Agent" as an example:

It can continuously track new papers in a research field;

Automatically sort out citation networks and methodological differences;

Reasonably infer trends and generate research summaries;

Even recommend potential collaborators according to your research interests.

This type of agent is not just a one - time "search tool" but an information analysis engine with long - term "cognitive memory" and dynamic "learning ability". They collaborate to form a continuously evolving knowledge network, greatly improving information screening and insight capabilities.

Communicational: Agents Can Communicate, Collaborate, and Negotiate with Each Other

Compared with the traditional human - centered Web, the real change in Agentic Web is that agents can collaborate with other agents, forming a multi - agent system similar to a "digital organization".

In the scientific research field, in a multinational research project, agents from different schools can:

Automatically synchronize experiment schedules;

Share data sets;

Generate joint results;

Automatically allocate authorship and funding ratios.

In the manufacturing or supply chain, agents from different enterprises can connect demands in real - time, respond to changes, and autonomously negotiate terms. This cross - agent collaborative workflow relies on a set of new communication protocols (such as MCP and A2A), supporting semantic alignment, task collaboration, and multi - party autonomy.

In short: the Web is no longer a bridge between humans and machines but a stage for agents to operate on.

V. Challenges: The Complex Problems and Future Bottlenecks of Agentic Web

Although Agentic Web shows an exciting prospect, to truly become the next - generation Internet in reality, it faces a set of systematic, intertwined, and interdisciplinary complex challenges —— it is far more than just improving the capabilities of AI agents. It also concerns the