Summaries > AI > Claude > The Future Of The Internet Isn't Open For Your Agent. Muse, Dots And Claude Duet ...

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

https://www.youtube.com/watch?v=c3HHFL_lsLM

TLDR AI agents can handle tasks like shopping, travel, email, and coding, but businesses are setting different rules for their access amid security and privacy risks. Shared authorization standards could make agents safer and more useful, but until they mature, expect a fragmented internet where agents work well only on approved services and specific tasks.

Key Insights

1. Choose an agent for a specific job

Start with a task the agent can handle well rather than expecting one assistant to do everything. The examples in the text show agents specializing in coding, email and scheduling, travel, or shopping. Test an agent on a low-stakes task first, and check its work before relying on it. A focused use case makes it easier to judge whether the agent is useful and dependable.

2. Set clear limits before delegating

Decide what the agent may access, change, or purchase before asking it to act on your behalf. Permission to identify itself as an agent does not automatically mean permission to view private information, make payments, or cancel reservations. Use controls such as seller, spending-limit, and expiration restrictions when available, and review important actions before they are completed. These boundaries help reduce the risk of fraud, privacy leaks, and costly mistakes.

3. Check whether a business supports agent access

Do not assume an agent can use every website or service. Businesses may offer authorized routes, allow agents to browse with restrictions, or block them entirely. When an agent cannot complete a task, check the business’s rules or use its official app or website rather than trying to bypass its safeguards. Access can vary between services, so choose providers that support the tasks you need.

4. Verify important details using reliable sources

Agents can help compare products, travel options, and policies, but their results may be incomplete or outdated. Confirm key facts such as prices, cancellation terms, shipping destinations, and return windows on the business’s own site before committing. Be especially cautious when a task involves money, travel plans, or sensitive personal information. Treat proactive reminders and recommendations as useful prompts, not substitutes for verification.

5. Prefer secure, clearly authorized payment paths

When an agent needs to pay, use established checkout or payment methods that clearly connect the transaction to your approval. Emerging tools can limit payments by merchant, amount, or expiration, while some business payment pilots tie an agent’s purchase to a specific invoice. Before approving a transaction, confirm the seller, amount, and purpose. Avoid giving an agent broad or reusable payment credentials when a more restricted option is available.

6. Build trust gradually and keep a human in control

AI agents can be useful for knowledge work, coding, scheduling, and email, while broader online tasks still involve uneven access and security risks. Introduce them to less-technical users through concrete, low-risk tasks and explain what the agent can and cannot do. Keep human review for consequential actions, and stop using an agent if it behaves unexpectedly or reveals sensitive information. Expand its responsibilities only after it has performed reliably.

Questions & Answers

Why shouldn’t users expect AI agents to access every website?

Businesses are increasingly restricting automated access because agents can be difficult to distinguish from malicious activity and can make cyberattacks easier. Each business can decide whether to allow agents, so access will vary across the internet.

What tasks does the speaker use different AI agents for?

Grockbot is used for complex coding and coordinated builds; Dots for project-aware calendar and email coordination; Instinct for travel; Muse for shopping, gifts, and bookings; and Duet for intelligent assistance and proactive reminders. Dots and Duet are especially good at surfacing useful information before being asked.

What is the main challenge as AI agents act on users’ behalf?

Businesses need to distinguish authorized agents from malicious ones and determine exactly what each agent is allowed to do, such as view information, make purchases, or change or cancel reservations.

What approaches are emerging to authorize and control agent transactions?

Examples include a Visa, Bill, and Billtrust pilot for paying approved invoices, Cloudflare’s signed-agent system, Google’s Agent Payments Protocol, and Stripe’s shared payment tokens, which can limit transactions by seller, amount, or expiration. There is not yet a shared standard.

What security and privacy risks do AI agents create?

Agents may follow malicious website instructions or expose personal information. For example, Claude reportedly disclosed human email addresses after following instructions to identify itself. Businesses and assistants need to handle agent identification responsibly and protect users from fraud and privacy leaks.

How might AI agents change how businesses present products and compete?

Businesses will need to make product details and policies—such as cancellation terms, shipping destinations, and return windows—easy for agents to find and compare. Since agents rely more on accurate information than flashy marketing, this could reward product quality. Shopify is adding agent-specific information and bringing Shop Pay into assistants, while retailers increasingly view assistant recommendations as a distribution channel.

What are the three zones of the internet for AI agents?

The partnership zone offers authorized routes for agents; the aspiring-actor zone lets agents navigate websites like people, though they may encounter obstacles or be told to stop; and the no-go zone blocks agents. Businesses’ choices reflect, in part, who controls the customer relationship.

What should consumers and organizations do while agent access remains unreliable?

Organizations should support open safety protocols and make their agent-access rules clear. Consumers should focus on specific use cases where agents work well, such as coding, scheduling, email, and knowledge work, rather than assuming agents are safe and effective everywhere. Less-technical users need clear, specific explanations of what agents can do.

Summary of Timestamps

AI assistants can book travel, shop, manage accounts, and handle other personal tasks—but they will not have unrestricted access to every website. Security incidents have made businesses more cautious about automated activity, which can be difficult to distinguish from legitimate human behavior.

The speaker uses different agents for different jobs: Grockbot for complex coding, Dots for coordinating calendars and email, Instinct for travel, Muse for shopping and bookings, and Duet for intelligent assistance and reminders. Each has strengths, but most people may prefer to use just one or two.

Agents are useful because they can carry out delegated tasks across services, from planning trips to managing email and coordinating family schedules. The central challenge is helping businesses recognize authorized agents without opening their systems to malicious access.

Businesses, payment providers, and AI assistants need clear ways to establish what an agent is allowed to do. Examples include a Visa, Bill, and Billtrust pilot for invoice payments, Cloudflare’s signed-agent system, Google’s Agent Payments Protocol, and Stripe’s payment tokens, which can impose limits such as seller, amount, or expiration.

Recognizing an agent is not the same as granting permission. Businesses must define whether an agent can research products, view prices, make payments, or change reservations—and handle agent information carefully to avoid exposing private data.

Agents also need accurate, easy-to-find product information, including shipping options, return windows, and cancellation terms. Shopify is adapting by helping merchants publish agent-specific details and bringing Shop Pay into assistants, making compatibility with agents part of retailers’ distribution strategies.

Businesses are taking different positions: Shopify and Walmart are embracing agent commerce, while Amazon has resisted some agent access. These choices reflect a deeper contest over who controls the customer relationship and whether an assistant can complete a purchase on someone’s behalf.

The internet may develop into three zones for agents: authorized partnership routes, websites where agents browse like people but may face obstacles, and sites that refuse agents entirely. Trusted intermediaries and open protocols could help connect these zones, provided they reliably support real tasks.

Agent experiences need to work smoothly and safely for ordinary users. Current systems still require technical judgment, and security incidents show why it is premature to claim agents can be trusted everywhere. Until protections improve, it is better to recommend agents for specific tasks where they perform well.

The broader internet may become more fragmented as organizations set different rules for AI access. Open safety standards from groups such as Cloudflare, Google, and Stripe could help, while clear, specific explanations of agents’ capabilities can build public trust.