How AI Agents Could Improve the Internet—and Make It Worse for Everyone Else

The internet was built around a relatively simple distinction: people browse, while bots automate limited tasks. That distinction is breaking down.

Personal AI agents can now act online for individual users. They may search for travel options, negotiate prices, or attempt to book flights and hotel rooms. Unlike conventional crawlers, these systems are not merely collecting information for a database. They are making decisions and taking actions on a person’s behalf.

That creates a problem the web’s existing systems were not designed to solve. Websites must decide whether a fast-moving visitor is a malicious bot, an authorized automated service, or a legitimate agent working for a customer. Many currently treat all machine-driven activity as a security risk.

The result could be an internet that becomes more convenient for people who use AI agents—but more restrictive, expensive, and difficult to navigate for everyone else.

AI agents are a new kind of web visitor

Traditional web traffic has generally fallen into two broad groups:

• Human users, who browse at human speed and interact directly with websites.
• Bots, which perform automated tasks such as indexing pages or monitoring content.

AI agents do not fit neatly into either category. They are software, but they may be acting on behalf of a specific person with a legitimate request. Their purpose is not necessarily to scrape or manipulate a website. An agent might be trying to compare hotel rooms, find a flight, or negotiate a purchase according to instructions supplied by its user.

The difference is that an agent can perform these actions much faster and more continuously than a human. It may send many requests, open numerous pages, and interact with several services in a short period.

For a website, however, speed and automation can look much like abuse. A system designed to prevent scraping, fraud, denial-of-service attacks, or automated account creation may block an agent before it has a chance to establish that it is acting for a real customer.

That is why AI agents represent more than an incremental increase in bot traffic. They create a third category of web visitor: automated software with a human principal and a potentially legitimate purpose.

The early encounter is already going badly

The supplied reporting describes a rapid increase in agent activity. Apps such as OpenAI’s Dots and Meta’s Muse have helped bring personal AI agents to a broad consumer audience. Muse was reportedly downloaded five million times in its first month, while agentic web traffic tripled from August to September.

Those figures indicate momentum, but they do not by themselves establish how much useful activity agents are completing or how the traffic was measured. The underlying trend is clearer: more consumers are asking software to perform tasks online, and websites are responding defensively.

Many sites have blocked agents that users sent to negotiate prices or plan vacations. From the site operator’s perspective, that reaction is understandable. An unknown agent can consume computing resources, probe systems at high speed, distort demand signals, or attempt transactions without behaving like a normal customer.

From the user’s perspective, the same block can feel arbitrary. The agent may be carrying out a perfectly ordinary request, such as searching for the best available room or comparing travel options. The website’s defenses cannot reliably distinguish that activity from hostile automation.

The conflict is therefore not simply between people and machines. It is between web infrastructure built for human interaction and software that is trying to represent humans at machine speed.

Why the web may be redesigned around agents

If agent use continues to grow, websites will face pressure to accommodate it. That could lead to a more structured internet in which services provide clearer ways for authorized agents to search, compare, and transact.

Such a shift could have genuine benefits:

• Agents could handle repetitive searches and form-filling.
• People could compare services without visiting dozens of websites manually.
• Websites could create controlled channels for automated purchasing or negotiation.
• Businesses might receive more clearly identified requests instead of undifferentiated bot traffic.

But the design choices will determine who benefits. A system optimized for agents may not remain equally accessible to people using ordinary browsers.

For example, websites could prioritize machine-readable interfaces over visual pages, or place more functions behind authenticated services. That may make transactions faster for recognized agents while adding friction for users who browse manually. A website could also impose stricter identity checks, account requirements, or access limits because it cannot distinguish between a trusted agent and a malicious automated program.

The danger is not that websites become more automated in itself. The danger is that automation becomes the assumed primary customer, while human access is treated as an exception.

The central trade-off: access versus control

Website operators have legitimate reasons to limit automated traffic. A system that accepts unlimited machine-speed requests can be exploited, overloaded, or used to manipulate markets and availability. Businesses also need to protect customer data and prevent unauthorized transactions.

AI agents create additional questions:

• Who is responsible when an agent makes a mistake?
• How can a website verify that an agent is acting with the user’s permission?
• Should agents be allowed to negotiate different prices?
• How can a service prevent one user’s agent from gaining an unfair advantage through speed?
• What happens when agents compete with one another for scarce flights, hotel rooms, or other services?

The source material does not resolve these questions, and there is no indication that a single solution has emerged. Blocking all automated visitors is simple, but it also excludes legitimate users. Allowing all agents unrestricted access would create obvious security and fairness problems.

A workable system would likely need to distinguish among identity, authorization, purpose, and behavior, rather than relying on the blunt category of “human” versus “bot.” Whether the industry can build that system without creating new barriers for ordinary users remains unsettled.

What a better agent-friendly web would need

An internet that supports AI agents without abandoning human users would need safeguards in both directions.

Clear identification

Agents should be distinguishable from anonymous or hostile automation, while users should not be forced to surrender unnecessary personal information merely to browse. Identification alone would not prove that an agent is trustworthy, but it could give websites more useful context than a binary bot check.

Explicit user authorization

An agent should operate within limits set by its user. Those limits might cover spending, the kinds of transactions allowed, or whether the agent may negotiate on the user’s behalf. Without meaningful controls, an agent could turn a simple instruction into an unwanted purchase or commitment.

Human-readable access

Machine-friendly systems should supplement, not automatically replace, ordinary websites. People need a way to inspect options, understand terms, and take over when an agent encounters an unusual or consequential decision.

Fair treatment of automated demand

If agents can act faster than people, they could gain an advantage in situations involving limited availability. Websites may need rules that prevent automation from overwhelming services or excluding customers who choose to transact manually. The source material does not specify what those rules should be, but the fairness issue follows directly from the speed difference between agents and people.

Accountability

When an agent misrepresents a user’s intentions, violates a site’s rules, or causes a financial loss, responsibility must be traceable. A web where automated systems can act without clear accountability would be difficult for both consumers and businesses to trust.

These are design requirements, not established features of the current internet. The transition is still developing, and the reporting available here describes the conflict more clearly than it describes the industry’s eventual solution.

How users should approach AI agents now

Consumers considering personal AI agents should treat them as delegated decision-makers, not as passive search tools.

Useful precautions include:

• Set narrow instructions. Specify what the agent may search for, purchase, or change.
• Use spending and transaction limits where the service provides them.
• Review important decisions yourself. Travel bookings, purchases, and cancellations can have consequences that an agent may not understand.
• Expect blocked sites. A failure may reflect a website’s security policy rather than a problem with the user’s request.
• Avoid assuming that speed equals a better deal. An agent can search quickly, but the supplied material does not establish that it will consistently secure lower prices or better terms.
• Check what information the agent can access. Greater convenience may require access to accounts, payment details, or personal preferences.

The most reliable role for an agent is likely to be one where its authority is limited and its actions are visible to the person it represents.

The choice ahead

AI agents could make the internet easier to use by handling tedious tasks that currently require repeated searches, forms, and transactions. But they also expose weaknesses in systems that were built around a sharp divide between people and bots.

If websites respond by blocking all automated activity, legitimate agent use will remain frustrating and fragmented. If they respond by redesigning everything around machine visitors, human users may face more authentication, less transparency, and reduced control over how services work.

The outcome will depend on whether the next generation of web infrastructure treats people as the principals behind agents—or as obstacles to a faster automated marketplace. A better internet for AI agents will only be better for everyone else if human access, accountability, privacy, and fair participation remain part of the design.

Sources

• “Making the internet better for AI agents might make it worse for everyone else,” published October 19, 2026. https://www.technologyreview.com/2026/10/19/1147186/making-the-internet-better-for-ai-agents-might-make-it-worse-for-everyone-else/