The Agentic Web Is Here: From the Semantic Web Vision to an Internet of AI Agents

From Semantic Web to Agentic Web: A Vision Two Decades in the Making

In 2001, a decade after inventing the World Wide Web, Tim Berners-Lee and two colleagues published an article in Scientific American proposing the idea of the "Semantic Web": an internet navigated on our behalf by autonomous software agents capable of handling our various needs. The vision seemed ahead of its time—web pages were primarily made for humans, and machines struggled to understand their meaning.

More than twenty years later, with large language models and tool-calling capabilities maturing, this vision is returning to the center of discussion in the form of the "agentic web." Both share the same core proposition: the internet is no longer merely an information space for people to browse, but an execution environment where software agents can operate autonomously and complete tasks on our behalf.

Semantic Web vs. Agentic Web: Same Origin, Different Paths

The Logic of the Semantic Web: Letting Machines Read Web Pages

Berners-Lee's approach was to add machine-readable semantic annotations to web pages, using structured metadata to let software agents understand the meaning and relationships of content. The key to this route was "description"—first model the world into a machine-understandable knowledge graph, then have agents act based on that knowledge. It depended on broad cooperation from content producers, making implementation very difficult.

The Semantic Web's technical approach used the Resource Description Framework (RDF) as its data model, expressing relationships between entities as triples (subject–predicate–object), and employed the ontology language OWL to define domain concepts and constraints, enabling data from different sources to be interlinked into a unified knowledge graph. At the web page level, structured annotations (such as the later schema.org vocabulary and JSON-LD format) let machines identify information like prices, authors, and events on a page. This system also gave rise to the Linked Data movement, which emphasized identifying entities with unified URIs and linking them to one another. Its fundamental limitation was that the annotation cost fell on each content provider while the benefits were diffuse and delayed, providing insufficient incentive to drive network-wide adoption.

The Logic of the Agentic Web: Letting Machines Operate Web Pages Directly

Today's agentic web takes a different path. Agents don't need websites to proactively provide semantic annotations; instead, they complete tasks by understanding natural language instructions, calling tools and APIs, and directly operating existing interfaces. In other words, it bypasses the precondition of "semanticizing the entire web" and shifts the cost of understanding from the content side to the model side.

The significance of this shift: the Semantic Web required the entire internet to first transform itself, whereas the agentic web has agents adapt to an internet that has not yet been transformed. The former is a top-down infrastructure project; the latter is bottom-up capability penetration.

Tool use / function calling refers to a model outputting structured call requests during generation, which an external runtime executes as real APIs or functions, then returns the results to the model to continue reasoning. It extends the model from "only generating text" to an executing entity capable of querying databases, sending emails, and submitting forms. Compared with directly operating graphical interfaces, tool calling is more stable and efficient, but requires service providers to offer usable interfaces; browser automation (such as page parsing based on the accessibility tree and actions like clicking and typing) serves as a fallback, covering websites that don't expose APIs. Together, the two form the main channels through which agents act on the existing internet.

Why This Vision Is Feasible Now

The conditions supporting the agentic web's implementation have gradually fallen into place over the past few years. Models have gained multi-step reasoning and tool-calling capabilities, able to break vague user intent into executable sequences of operations; browser automation and the API ecosystem give agents genuine "hands-on" channels; and collaboration protocols between agents are gradually taking shape, allowing tasks to flow among multiple agents.

The autonomous agents Berners-Lee envisioned back then, constrained by the technical conditions of the time, could only remain at the conceptual level. Today's tech stack, for the first time, takes "agents navigating the internet on our behalf" from papers to runnable systems. This is the basis for the judgment that "the agentic web is here."

Inter-agent collaboration protocols aim to solve the problem of "how one agent discovers, delegates to, and settles with another agent." Typical designs include: capability description documents that declare the tasks an agent can perform and the input/output formats, enabling callers to match automatically; standardized message-passing mechanisms to complete multi-turn task handoffs; and identity authentication and payment credentials so cross-organizational calls can be attributed and billed. The value of such protocols lies in extending a single agent into a division-of-labor network—one retrieves, one executes, one verifies—but they still need to address the lack of unified interoperability standards and security issues such as malicious agents making impersonated calls.

The Real Problem Isn't Technology—It's Trust and Governance

The truly thorny part of the agentic web is precisely the question that the Semantic Web discussion could not avoid either: when software agents act online on behalf of users, who is responsible for their behavior?

Agents need access to users' accounts, payment capabilities, and personal data, which means permission boundaries must be strictly defined. Whether every operation an agent performs is auditable and traceable determines whether users are willing to entrust it with real tasks. Moreover, when large numbers of agents access websites on behalf of humans, traffic structure, the content ecosystem, and even advertising models will be affected—how websites should distinguish human visitors from agent visitors has no mature answer yet.

The Semantic Web era's failure partly stemmed from a lack of incentives: content providers had no motivation to add extra annotations for machines. The agentic web faces a different kind of incentive mismatch: websites may not be willing to be operated by agents, while agents need access to those websites. This tension will determine how far the agentic web can go.

What It Means for Developers and Product Teams

For those building products, the changes brought by the agentic web are structural. Past product design treated human users as the sole operating subject, with interfaces, flows, and conversion paths all centered on human attention. When agents become a new class of "users," products need to be designed for two types of operators at once: interfaces humans can understand and interfaces agents can reliably call.

This also explains why open APIs, structured data, and clear permission models have become important again—they are not just engineering norms but preconditions for whether agents can integrate smoothly. The machine readability the Semantic Web sought is being needed again in a more pragmatic way.

From a magazine article in 2001 to runnable agent systems today, the internet's form is undergoing a substantive transformation. The agentic web is not a simple revival of the Semantic Web but another implementation path toward the same goal under new technological conditions. Whether it can fulfill Berners-Lee's original vision depends on whether trust mechanisms and ecosystem collaboration beyond technology can keep pace.

One-line Summary

From Semantic Web to agentic web, the same vision implemented along a different path, with the bottleneck shifting to trust and governance.

Paragraph Summary

The article revisits Tim Berners-Lee's 2001 Semantic Web vision and compares it with today's emerging agentic web. The Semantic Web required content providers to proactively offer machine-readable annotations, a top-down infrastructure route that struggled to gain traction due to lack of incentives; the agentic web instead has agents directly adapt to the existing internet through natural language understanding, tool calling, and interface operation, shifting the cost of understanding from the content side to the model side. The maturity of multi-step reasoning, browser automation, and the API ecosystem makes this vision feasible for the first time. But the real obstacles lie beyond technology: agents' permission boundaries, operation auditability, how websites distinguish human from agent visitors, and the incentive mismatch between content providers and agents will all determine its ceiling. The article concludes that product teams need to design for both human users and agents as operators.