From Multi-Agent Systems and the Semantic Web to Agentic AI: A Unified Narrative of the Web of Agents
This paper presents a unified narrative of the "Web of Agents" evolution from 1990 to 2026, arguing that a "semantic-effort migration" from platform coordination to data annotation and finally to model interpretation defines three distinct generations, while offering a comparative framework, system classifications, and an analysis of recent institutional convergence to identify persistent challenges and future directions for Agentic AI.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine the internet not as a library of static books (documents), but as a bustling city filled with autonomous workers (agents) who can talk to each other, make deals, and get things done on your behalf. This paper calls that future city the "Web of Agents."
The authors, a team from the University of Luxembourg, argue that we have been trying to build this city for 30 years. They trace the history of these attempts through three distinct "generations," showing how we kept moving the "brain" of the system to different places, only to realize that each move created a new problem.
Here is the story of the Web of Agents, told in simple terms with some creative analogies.
The Three Generations: Moving the "Brain"
The paper argues that the history of agent technology is a story of where we decided to put the "meaning" (semantics) of the conversation.
Generation 1: The "Strict Club" (1995–2005)
- The Idea: Imagine a high-end country club where every member must wear a specific uniform and speak a very formal, rigid language (like Latin) to enter.
- How it worked: This was the era of Multi-Agent Systems (MAS). Researchers built complex platforms (like a club manager) that enforced strict rules. Agents had to follow a specific "speech act" protocol (like saying "I request" or "I promise" in a very specific way).
- The Problem: The internet is like a chaotic, open street market, not a country club. The "club rules" were too heavy and complex for the open web. The technology worked inside closed labs but failed to connect with the rest of the internet.
- The Lesson: You can't force the open web to act like a private club.
Generation 2: The "Annotated Library" (2001–2012)
- The Idea: If the agents can't agree on a language, let's just label the books in the library so clearly that anyone can understand them.
- How it worked: This was the Semantic Web. Instead of making the agents smart, researchers tried to make the data smart. They wanted to add complex tags (ontologies) to every webpage so machines could understand the meaning of the content.
- The Problem: It was too much work. Labeling the entire internet manually was like trying to write a dictionary for every word in every book in the world. It was too expensive, and the system was too brittle (if a label was slightly wrong, the whole thing broke).
- The Lesson: You can't rely on humans to manually label the entire internet.
Generation 3: The "Smart Intern" (2020s–Present)
- The Idea: Stop trying to label the books or force a specific language. Just hire a super-smart intern who can read anything, understand context, and figure things out on the fly.
- How it worked: This is the current era of LLM (Large Language Model) Agents. The "meaning" is now inside the AI model itself. The AI doesn't need pre-labeled data; it uses its training to understand natural language and figure out what tools to use.
- The Result: This finally worked! Because the AI is flexible, we can finally build a Web of Agents.
- The Catch: Because the AI is "guessing" the meaning based on patterns rather than following strict rules, we can't always prove why it did something. It's flexible, but it's not perfectly verifiable.
The New "City Infrastructure" (2024–2026)
The paper highlights a massive shift happening right now (late 2024 to 2026). For the first time, the "plumbing" is being installed to let these agents actually do business.
- The Protocols (The Phone Lines): New standards like MCP (Model Context Protocol) and A2A (Agent-to-Agent) are being launched.
- Analogy: Think of MCP as a universal power outlet. No matter what tool you have, you can plug it into the AI.
- Analogy: Think of A2A as a universal phone number system. It lets Agent A call Agent B, regardless of who made them.
- The Payment System (The Wallet): For decades, agents couldn't pay for things. Now, giants like Visa, Mastercard, and Stripe are building specific payment rails for agents.
- Analogy: Before, agents were like kids who could talk but had no money. Now, they have their own credit cards with spending limits.
- The Rules (The Police): Governments (like the EU) and safety groups are starting to write laws specifically for these agents, ensuring they don't go rogue.
The Seven "Lessons" and "Challenges"
The authors总结出 (summarize) seven key lessons from this 30-year journey. Here are the big ones:
- Everything Must Fit: You can't just have a great brain (AI) if the phone lines (protocols) or the payment system (economics) don't work. All parts must align.
- The "Brain" Migration: We moved the "brain" from the Platform (Generation 1) to the Data (Generation 2) to the Model (Generation 3). Each time we moved it, we gained flexibility but lost some safety guarantees.
- Money Matters: The Semantic Web failed because no one had a business model to pay for the work. Now, payment networks are stepping in to fix this.
- The "Trust" Trap: Every time we tried to find agents, we built a central "phone book" (a registry). These always fail at a massive scale. We need a decentralized way to find agents, like a peer-to-peer network.
- The "Black Box" Problem: Because modern AI figures things out on its own, we can't always verify if it's doing the right thing. The next step is to add "signed contracts" so we can verify the AI's actions without losing its flexibility.
- The "Consent" Puzzle: If Agent A asks Agent B to do something, and Agent B asks Agent C, who actually gave permission? We need better rules for passing permission along.
- Regulation is Catching Up: For the first time, laws are being written while the technology is being built, not years later.
The Future: A "Verified Contract" Era?
The paper predicts a fourth phase is coming. We are currently in the "flexible but unverified" phase. The future will likely be a mix: Flexible AI that is wrapped in Signed Contracts.
- Analogy: Imagine a very creative artist (the AI) who is allowed to paint whatever they want, but they must sign a legal document at the end of every painting proving they didn't break the rules. This gives us the best of both worlds: creativity and safety.
Summary
The paper concludes that the "Web of Agents" is no longer just a sci-fi dream. The technology (AI), the tools (protocols), the money (payments), and the rules (regulations) are finally coming together. The challenge now isn't building smarter agents; it's building a trustworthy, safe, and economically viable city where they can all live and work together.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.