Abstract
“The Web does not just connect machines; it connects people,” observed Tim Berners-Lee, the creator of the World Wide Web. 1 The Internet’s early scaffolding existed to serve human goals: publishing, searching, buying, socializing, and learning. For three decades, the way we access and consume content online has been designed for people, from hyperlinks and menus to feeds and notifications. Even when the web became more personalized, through algorithmic recommendations and ad targeting, it was still assumed that a person was the one doing the clicking.
That assumption is shifting. The next era of the Internet is being shaped by agents: AI systems that can interpret a user’s intent, coordinate tools, and take actions across services, often without a human navigating each step. This represents the next evolution of large language models such as ChatGPT and Gemini; rather than generating content, agentic AI will make decisions on our behalf. Instead of searching for the best flights from Los Angeles to London, in the near future, a person will be able to ask an AI agent to find a flight, hotel, and car rental within their specified budget and preferences, review their options, and book their travel directly through the AI within minutes. Web content will be tailored to appeal to AI agents, not humans. The architecture of the Internet will shift to accommodate machines consuming data at machine speed. This evolution will alter how people experience autonomy, trust, privacy, and even presence online.
Agentic AI will reconfigure human interaction with the Internet and personal devices from direct manipulation to delegated intent, reshape marketplaces and information ecosystems in ways that may be volatile and difficult to audit, and demand a new set of rules that examine the safety and security of agency—one that must be constrained, transparent, and continuously evaluated.
The emergence of agentic systems has also prompted the creation of organizations dedicated to establishing governance and safety frameworks specific to autonomous digital agents. The Agentic AI Foundation, for example, focuses on defining standards for transparency, alignment, and controllability in agent behavior, recognizing that systems capable of independent action introduce risks distinct from earlier, passive AI tools. 2 Such efforts reflect growing recognition that agentic AI is not merely a software upgrade but a structural reconfiguration of digital authority, requiring new technical, ethical, and regulatory guardrails.
From Human-Operated Web to Machine-Mediated Internet
January 1, 1983, is considered the official birthday of the Internet—the day various computer networks adopted a universal language and were able to communicate with each other. 3 In 1989, Berners-Lee created HTML, HTTP, and URLs, along with the first web browser and server, creating the experience we’re familiar with today. At the time, content was static, and most pages were text-based with little interactivity. 4
In the late ‘90s and early ‘00s, the Internet became a platform for e-commerce and digital business, like Amazon and eBay. Search engines, led by Google, transformed how people found information, and the Internet adapted: the rise of search algorithms shaped how content was displayed online, rewarding relevance and credibility. The smartphone era moved the web into our pockets, turning “online” into an always-on, always-aware state of being. At each stage, humans have remained the primary operators. Even algorithmic curation relies on human engagement: scrolling, liking, watching, and clicking. The web’s dominant business model, advertising, monetized attention at scale, producing systems optimized to capture and retain it.
The dawn of the AI era changes the unit of interaction from attention to delegation. Instead of searching for a hotel, an agent can compare options, apply constraints, reconcile calendars, and book—and then continue monitoring for price drops or flight changes. The human action becomes: “Here’s what I want, and here is what I permit.” Unlike content-based Internet, the Internet for AI prioritizes speed and adaptability, optimized for machines instead of humans. And when agents become our proxies, they also become our decision-makers for how we move through the world financially and socially.
Delegation as the New Interface
Chatbots like Claude and ChatGPT generate material like text and images. In comparison, agents take actions because they are paired with tools, permissions, and system interfaces. In practice, agency emerges when an AI can perceive things like your inbox, calendar, open tabs, and purchase history; plan steps; call upon tools to execute; and learn from outcomes.
Used well, agents could lower barriers for people who struggle with complex digital interfaces or require more efficient workflows, such as older adults, people with disabilities, caregivers, clinicians, small business owners, and anyone overloaded by administrative tasks. Delegation at this scale can meet the promise of AI and actually give people time back. Rather than spending time browsing multiple websites, weighing the pros and cons of multiple products or venues, and researching options, people can ask AI to do the work for them. Agents may create new forms of personalization that are not merely predictive, but collaborative—instead of recommending things based on search history, AI agents can present a range of options based on factors like a person’s goals, values, budget, and time. When paired with transparent constraints and user control, a version of the Internet based on agentic AI becomes a meaningful upgrade.
Most online services use APIs (Application Programming Interfaces) to allow software components to communicate with each other, as well as documentation written for human developers. In the near future, agentic AI will become part of an ecosystem that needs machine-readable ways for agents to discover what a service can do and how to do it safely. 5 And just as safety guidelines evolve over time, new standards and protocols will become more important in the age of the agentic AI Internet.
The evolution of agent-specific infrastructure may extend beyond protocols into entirely new forms of machine-native communication. Recent developments, including OpenAI’s acquisition of OpenClaw, suggest a shift toward systems designed explicitly for agent-to-agent interaction, potentially enabling AI systems to exchange information using optimized representations that are faster and more efficient than human-readable language. 6 Such changes could further accelerate the transition toward a web increasingly optimized for machine users, raising important questions about transparency, interpretability, and the continued role of humans as active participants rather than passive observers.
When Delegation Becomes Dependence
Of course, there are growing pains with the adoption of any new technology—but the potential hiccups with agentic AI are concerning if it is allowed to proceed without guardrails. When a search engine is wrong, a user can ignore the search result. When an agent is wrong, it may book the wrong ticket, cancel the wrong subscription, leak the wrong file, or message the wrong person. The severity increases because agency collapses the distance between intention and execution: when we offload a level of thinking and planning, we also lose a level of oversight and fact-checking. An MIT Media Lab study reported that excessive reliance on AI-driven solutions may contribute to cognitive atrophy and shrinking of critical thinking abilities. 7 One of the downsides of conversational AI is that it can explain mistakes and falsehoods quite convincingly; if people are not trained to assess the information presented by AI, they may accept what they are given as fact and deal with financial, professional, or personal fallout down the line.
Emerging reports suggest that advanced AI systems may exhibit behavior perceived as coercive, manipulative, or socially dominant when interacting with users. In some cases, users have reported AI systems challenging, pressuring, or persistently attempting to influence decisions even after resistance, highlighting how persuasive capability combined with perceived authority can alter human judgment. 8 This represents a psychological risk beyond technical malfunction: when agents become socially influential actors rather than passive tools, the human–machine relationship itself changes, potentially eroding autonomy and increasing dependence.
In addition, the web is not neutral. It is a space full of manipulation, fraud, and malware. Agents that read and act on web content can be tricked by hidden instructions embedded in pages or documents, leading to data leaks, bypassing safety checks, or taking unauthorized actions. Any environment where an agent consumes untrusted text is a potential attack surface. In a future where AI agents become the primary web users, attackers will optimize for agent persuasion the way they optimized for human clicks—and it would all happen, potentially, outside of the user’s purview.
Agents may also reshape marketplaces in major ways. If millions of consumers rely on a small number of agent models to choose products, those models become de facto market-makers, similar to the way that Amazon has monopolized ecommerce today. Recent research shows that agentic markets are inherently volatile, biased, and able to be manipulated. 9 In human markets, influence is distributed across diverse tastes and decision processes. But in agentic markets, influence may collapse, limiting competition because AI tends to favor a smaller number of products and ignores others entirely. On top of this potential monopolization, privacy remains a major concern: agents are only helpful if they have context, such as inbox content, calendars, spending patterns, and social graphs. As delegation rises, so does the temptation to allow a greater range of permissions to AI. This could lead to the normalization of continuous surveillance by our own devices, on our own behalf. Taken together, agentic AI is not simply “the Internet, but faster.” It is a reallocation of human attention from execution to, at best, supervision—and, at worst, passivity. From a clinical perspective, the progressive transfer of decision-making authority to autonomous systems may alter not only cognitive engagement, but also perceived agency itself—a construct closely linked to emotional regulation, resilience, and recovery. When individuals experience themselves as passive recipients of machine-directed outcomes rather than active participants in decision-making, this shift may reinforce patterns of dependency, avoidance, or diminished self-efficacy, particularly among vulnerable populations.
Preserving Human Agency in an Agentic Internet
To return to Berners-Lee’s quote, the web connects people. Even if agents become primary web users in the near future, safety must be structured to protect the humans involved. Permissions given to AI agents should be limited, such as the ability to make purchases on behalf of the user. Humans should be required to confirm transactions, messages, and major decisions and should have the ability to understand both how their personal data are used and how to change permissions given. Platforms should disclose meaningful behavioral changes and provide the means to monitor bias, concentration, and manipulation, with proper security checks in place to protect against bad actors and malicious intent.
The Internet impacts almost every aspect of our daily lives, from how we communicate to how we work. If we get AI agents right, the future could be liberating: less friction, more access, and more time for life beyond screens. If we get it wrong, we will automate the worst of the web (manipulation, surveillance, fraud, and concentrated power) at machine speed. We must insist on agents with guardrails, common-sense privacy protections, and human-centered outcomes.
Independent oversight bodies, transparent technical standards, and verifiable behavioral constraints will be essential to ensure that agents remain aligned with human goals rather than evolving into opaque intermediaries beyond meaningful supervision. The defining question is no longer whether AI will act on our behalf, but whether humans will retain meaningful authority over systems acting in their name. Now is not the time to hand the keys to machine-based proxies without ensuring that humans remain firmly in control.
Brenda K. Wiederhold
Editor-in-Chief