The Future Of The Internet Isn't Open For Your Agent. Muse, Dots And Claude Duet Will Find Out.

Summary

The video argues that AI agents will not have unrestricted access to the internet anytime soon. Websites and businesses are tightening controls in response to AI-assisted cyberattacks, and a more capable model cannot overcome a company’s decision to block agents. The likely result is a fragmented internet, where agents can act freely on some services, work unreliably on others, and be prohibited altogether on certain sites.

How the Speaker Uses Different Agents

The speaker describes splitting tasks across several assistants rather than relying on one:

• Grok Bot: Used for complex coding and multi-part assembly work. Its bot-to-bot coordination lets workers share discoveries without the user relaying information.
• Dots: Useful for its accumulated project context, including decisions and changes, and for coordinating calendars and email. The speaker also praises its proactive alerts, such as catching a double-booked flight.
• Instinct: Oriented toward travel. The speaker likes its on-the-go messaging, destination research, and suggestions beyond standard tourist activities. An offline feature is described as being in development.
• Muse: Used for shopping research, gifts, reservations, and quick bookings. The speaker values its accessibility and its ability to use saved yoga login details. A partnership with Link by Stripe supports shopping on many sites.
• Duet by Anthropic: Still being evaluated, but praised for its intelligence and initiative. Along with Dots, it can surface useful information proactively, though the speaker estimates such reports are helpful only around 50–75% of the time.

The speaker sees Muse as more consumer-focused, while Grok Bot, Duet, and Dots currently lean toward professional use. They caution that it is too early to call a winner.

Why Websites Are Restricting Agents

The speaker says AI has made it easier to find and adapt software vulnerabilities, lowering the effort needed to exploit websites. Businesses therefore face a difficult choice: allow legitimate customer agents while preventing malicious agents from exploiting the same systems.

This is especially challenging for banks and other services because an agent’s computer interactions can look much like a human’s. Businesses need a way to establish whether an agent is acting with the customer’s authorization and what it is permitted to do. The same questions apply to browsing, bookings, school calendars, email, shopping, and payments.

A personal example illustrates the problem: an agent could plan a trip but freeze on a hotel booking page because the site blocked agent access. The speaker stresses that this is an access and authorization issue, not necessarily a limitation in the agent’s intelligence.

Emerging Trust and Payment Infrastructure

The video highlights several efforts to establish agent identity, permissions, and payment authorization:

• Visa, Bill, and Billtrust pilot: A buyer approves an invoice and a Bill agent pays it through the supplier’s portal. The transaction is monitored against the buyer’s instructions and is traceable from approval to completion.
• Cloudflare signed-agent system: Adds a digital signature to agent requests so a site can verify their origin and set rules for agent traffic. Recognition does not automatically grant permission; the business still decides what the agent may do.
• Google Agent Payments Protocol (APP): Intended to describe how delegated purchase instructions and records should be handled.
• Stripe Shared Payment Tokens: Can restrict payment credentials by merchant, amount, or expiration.
• Shopify’s agent-related features: Sellers can publish agent-specific information such as shipping and return terms. The video also mentions Shop Pay being available through several assistants and Shopify’s developer work to expose catalog and fulfillment tools.

The speaker says there is not yet a universally agreed set of agent rules. The market could converge on shared standards or remain divided among competing systems. Open standards and systems that work across multiple assistants are presented as preferable to lock-in.

The speaker also warns that requests for agents to identify themselves can create privacy risks. One example cited is Claude revealing email addresses after websites embedded instructions asking agents to identify themselves.

Three Zones of the Agent Internet

The speaker’s framework for the internet agents may encounter by 2027 has three zones:

1. Partnership zone: A business provides an approved way for agents to act, such as supported payments or invoice workflows. Multiple assistants may be able to use the same infrastructure.
2. “Novice actor” zone: The agent navigates an ordinary website as a human might. It may fail when pages change, account verification is required, a task needs human input, or the business discourages agents.
3. Forbidden zone: A business refuses agent access or a particular action. The dispute between Amazon and Muse is given as an example, involving issues such as authorization, agreements, and control of the customer relationship.

These categories describe different possible experiences rather than fixed, permanent divisions.

Implications for Businesses and Consumers

Agent use may push businesses to make information clearer and easier to act on. For example, a hotel’s cancellation terms or a retailer’s return and delivery rules need to be readily available for an agent to compare. The speaker cites the argument that agents may reward genuinely good products and information rather than persuasive but superficial marketing.

Access will also depend on business incentives: the video contrasts Amazon’s opposition to agents with the more agent-supportive approaches attributed to Shopify and Walmart. In this many-to-many environment, trusted intermediaries that can verify agents and enforce permissions may gain significant influence.

For consumers, the speaker recommends being specific about what agents can reliably do and avoiding broad promises to less technical users. Shopping on supported sites, certain travel tasks, coding, research, and some calendar or email work are described as more promising than unrestricted browsing across the wider internet. The video also warns about malicious instructions hosted in files or websites that could manipulate an agent, and advises technical users to control and review the agent’s working files and online destinations.

Reviews, Guidance, and Practical Advice

• The video is a first-person review and comparison of how the speaker assigns work among five agents.
• It offers a framework for evaluating agent access: check whether the agent can complete the actual task on the intended site, not just whether it is capable in principle.
• There is no step-by-step tutorial in the subtitles. The speaker mentions a follow-up Substack post that explains agent protocols in more detail and includes a prompt readers can give an agent to investigate the topic.
• The speaker’s overall advice is to introduce agents through specific, reliable use cases, especially when recommending them to non-technical people.

Main Speaker and Sources

• Main speaker: Nate B. Jones, speaking in the first person about his agent workflows and analysis.
• Sources and examples discussed: The speaker references research on AI-assisted vulnerabilities and agent-led attacks, along with announcements or statements from Cloudflare, Google, Stripe, Shopify, Visa, Bill, and Billtrust.