🕒 4 min read
Web of Agents: Will AI Agents Run the Internet?
The internet is changing. It’s not just pages anymore. There are agents now. These agents work for you. They book appointments. They compare prices. They write emails.
This idea isn’t new. It’s been around since 1990. But now it’s possible. Why? Because large language models got better. LLMs understand language. They get context. What was impossible before is now real.
A new paper explains this. Title: “From Multi-Agent Systems and the Semantic Web to Agentic AI”. Published on arXiv in 2026. Authors are from the University of Luxembourg. The main claim is simple: The Web of Agents went through three generations. In each generation, the semantic effort moved somewhere else.
Generation I: Everything Was on the Platform
In the 1990s, agents tried to talk. They used a language called FIPA-ACL. There was also KQML. Agents sent messages. But there was a problem. Everyone had to be on the same platform.
The platform set the rules. The intelligence was in the platform. Agents registered to a central directory. A broker would find them. Systems like JADE appeared.
What happened? It didn’t work. It was too rigid. It didn’t fit the real world. Companies didn’t adopt it. Platform-side coordination failed. The semantic effort was on the platform. That killed flexibility (Petrova et al., 2026).
Generation II: We Added Meaning to Data
In the 2000s, a new idea came. People said: “Let’s tag the data”. The Semantic Web was born. RDF was used. Ontologies were written in OWL. Queries ran on SPARQL.
This time, intelligence was in the data. Every page got labels. “This is a person.” “This is a price.” So machines could read it. Search engines like Swoogle were built. SPARQL endpoints opened.
Again, there was a problem. Who will tag everything? No one did. Data annotation was too expensive. Not even 1% of the web got tagged. The semantics-in-data approach didn’t scale. The effort was in the data. But the data was too big (Petrova et al., 2026).
3. Generation III: The Model Understands Everything
After 2020, LLMs arrived. The game changed. Now we don’t need to tag data. The platform isn’t required either. The model understands by itself.
GPT, Claude, and Llama read text. They figure out intent. They call APIs. This is “semantics-in-models”. The meaning is in the model weights.
Now agents talk in natural language. The MCP protocol came out. Anthropic announced it in November 2024. Google released A2A v1.0. Agents find each other. They talk. Discovery happens with vector search.
Visa and Mastercard joined too. Agents will make payments. Stripe is writing protocols. The Linux Foundation started the “Agentic AI Foundation”. The EU AI Act is active. From 2024 to 2026, everything is getting standardized.
What’s the problem with this generation? Hallucination. Security. The model can be wrong. The semantic effort is in the model. If the model fails, the agent fails (Petrova et al., 2026).
4. The Future: Hybrid Systems Will Win One generation isn’t enough. The future is hybrid. LLM + Knowledge Graph is combining. The model is flexible. The KG is accurate.
There are also computer-use agents. Claude looks at the screen. It clicks. OpenAI Operator does the same. You don’t need an API anymore. Agents use tools like humans.
So there are three rules:
Wherever the semantic effort goes, the problems come from there. It migrated from platform → data → model. Now it’s time for hybrids. WoA is not a document web anymore. It’s an action web. Agents will click for us. Read for us. Decide for us. But we need rules. That’s why the EU AI Act exists. The limits for agents are being drawn.
Are we ready? The model is ready. The protocols are ready. Payments are ready. The next step is trust.
Citations and References This blog post is based on the academic source below. All core claims, dates, and the three-generation model are taken from this work.
Primary Source Petrova, T., Bliznioukov, B., Puzikov, A., & State, R. (2026). From Multi-Agent Systems and the Semantic Web to Agentic AI: A Unified Narrative of the Web of Agents. arXiv:2507.10644v4 [cs.AI]. License: CC BY 4.0. University of Luxembourg, SEDAN – SnT.
In-Text Citations The three-generation semantic migration model was first defined by Petrova et al. (2026). The authors split the 1990–2026 period into Generation I: Platform-side coordination, Generation II: Data-side annotation, and Generation III: Model-side interpretation. The “semantics-in-data → semantics-in-models” shift is the central thesis of the study.
The MCP launch in November 2024, MCP specification in November 2025, A2A v1.0, Visa/Mastercard/Stripe protocols, and the phased rollout of the EU AI Act are all documented in the same paper under the “2024–2026 institutional convergence” section (Petrova et al., 2026).
The study also classified 16 representative systems using a four-dimensional framework: semantic foundation, communication paradigm, locus of intelligence, and discovery mechanism. Hybrid LLM-KG and computer-use agents were included in this classification.
Access: https://arxiv.org/abs/2507.10644 License: CC BY 4.0 – Free to share with attribution.
