Quick answer
AI agent internet access is becoming uneven, not open. In his latest video, Nate B Jones argues that agents such as ChatGPT Dots, Meta Muse and Claude Duet will meet three kinds of website in 2027. In the partnership zone, the business offers an approved route, such as a supported checkout. In the aspiring actor zone, the agent clicks through the site like a person and may get stuck. In the no-go zone, the business refuses agents outright. After the October wave of AI-assisted attacks, many sites have good reason to tighten up. A smarter model will not open a door a business has chosen to close.
Key takeaways
- Three zones: partnership (approved routes), aspiring actor (the agent tries the normal website) and no-go (agents refused).
- Security is driving the split. After “Black October”, sites have fresh reasons to be stricter with automated visitors.
- The core problem is authorisation: an agent’s clicks look much like a human’s, so a site struggles to tell your helper from an attacker.
- A trust layer is forming: Cloudflare signed agents, Google’s AP2 mandates, Stripe shared payment tokens and a Visa, BILL and Billtrust invoice pilot.
- Clear facts beat glossy pages. Agents need return windows, shipping and cancellation terms they can actually find.
- Retailers are choosing sides: Shopify and Walmart are saying yes to agents, while Amazon is saying no.
- Practical lesson: use agents where they reliably work today, and set limits before they spend money.
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Set Spending Limits Before an AI Agent Shops for You — cap amounts, limit shops and add approval steps before your agent can pay for anything
- How to Recommend an AI Agent to a Non-Technical Friend — match an agent to the jobs someone actually cares about, test it first and set safe defaults
In this guide
- Watch: the source video
- What are the three zones of AI agent internet access?
- Why are websites getting stricter with AI agents?
- Why can’t a website tell your agent from an attacker?
- How are payments and identity being built for agents?
- What do businesses now need to show your agent?
- Which AI agents does Nate B Jones use, and for what?
- What does the video not confirm?
- How should you use AI agents on the web right now?
- Who does this matter to most?
- Frequently asked questions
- The bottom line
- Related posts
More AI assistants than ever want to book your flights, buy your gifts and manage your accounts. The catch is that the websites they need are becoming more careful about who, or what, they let in.
This article breaks down Nate B Jones’s view of AI agent internet access: the three zones your agent will meet, why security is pushing sites to close doors, the payment and identity tools being built to fix it, and what it means for you as a user or a business owner.
Watch: the source video
What this video covers: Nate B Jones explains why the internet will not be open to personal AI agents any time soon, shares which agents he uses for which jobs, walks through new agent identity and payment schemes, and offers a three-zone framework (partnership, aspiring actor, no-go) for the web your agent will meet in 2027.
Video: Nate B Jones (@NateBJones, AI News & Strategy Daily). This summary and the opinions in it are our own; the agent setups, estimates and announcements are as described in the video.
What are the three zones of AI agent internet access?
Nate’s framework splits the web your agent will meet into three zones. They are not fixed categories, but they describe experiences most of us will start to have.
| Zone | What happens | Example from the video |
|---|---|---|
| Partnership | The business offers an approved route for agents to do the job | Shop Pay checkout inside assistants; the Visa, BILL and Billtrust invoice pilot |
| Aspiring actor | The agent works through the normal website as a person would, and may stall | Claude Duet’s browser on sites with no connector |
| No-go | The business refuses the agent or the action | Amazon’s dispute with Meta Muse |
The partnership zone is the most reliable. Several competing assistants can share the same approved route, so picking one agent does not lock you out of those businesses.
The aspiring actor zone is the messy middle. Your agent might finish the task. It might also hit a changed page, an account check or a step that needs you. If a site politely asks agents not to proceed, a well-behaved agent will often comply.
The no-go zone is about who owns the customer relationship. Nate says Amazon’s objections to Muse have included undisclosed agent access, no agreement in place and how login details are handled.
💡 Tip: When an agent task fails, ask which zone you were in. A no-go site will not change because you switch to a cleverer model. Try a business that offers a supported route instead.
Why are websites getting stricter with AI agents?
Because AI has made breaking into websites much easier, and businesses are responding. Nate links this to “Black October”, the run of security incidents and research in mid to late October 2026.
He points to research describing AI helping people develop and adapt working exploits across several applications in hours. Part of the web’s old protection was that finding and using flaws was slow and difficult. AI has pushed that difficulty much lower, for attackers as well as researchers.
He also mentions an investigation showing attackers using agents to coordinate, share findings and keep trying. So businesses now face two kinds of persistent software at once: agents doing legitimate work, and agents trying to break in.
⚠️ Warning: If you run a website, this is a good moment to review your own security basics before you think about welcoming agents. Our earlier piece on Black October and AI keys covers the first steps.
Why can’t a website tell your agent from an attacker?
Because agent clicks and human clicks look almost the same. Nate calls this a first-principles problem: computer use by an agent is, by design, similar to computer use by a person.
He cites reports that US banks have found AI agents inside customer accounts and are unsure how to respond. The bank cannot easily tell whether it is an enthusiast using an agent on their own account, or a bad actor using one to get in.
Scale makes it harder. If lots of people send agents to the same school portal or booking page at once, it can look like an attack. Nate notes that this has already happened with restaurant bookings, where people using agents to hunt for tables were reportedly banned from Resy.
He frames the fix as three decisions that must agree:
- The business needs to know it is dealing with an authorised customer.
- The payment provider needs to know what is allowed to be charged.
- The AI assistant needs to protect your account from attempts to redirect it.
Letting an agent near your card? Read our step-by-step guide on how to set spending limits before an AI agent shops for you, so every purchase stays within rules you chose.
How are payments and identity being built for agents?
Several companies are building ways for agents to prove who they are and what they are allowed to do. Nate describes them as smart people proposing guardrails, with no single agreed standard yet.
| Scheme | What it does (as described) |
|---|---|
| Cloudflare signed agents | The service sending the request attaches a digital signature that Cloudflare can verify, and the website sets rules for agent traffic |
| Google AP2 (Agent Payments Protocol) | Uses signed “mandates” as tamper-proof proof of a user’s instructions for a purchase |
| Stripe shared payment tokens | Payment credentials that carry limits, such as one seller, a maximum amount and an expiry (public preview) |
| Visa, BILL and Billtrust pilot | A buyer approves a supplier’s bill, a registered BILL agent pays it through the supplier’s portal, and the payment stays tied to the invoice (participating companies only) |
Recognition is only the first step. Nate’s example: a hotel might let a verified agent read a price and see a reservation, but not cancel it. Each permission is a separate business decision.
He also flags a side effect. According to the video, a reported GitHub issue described Claude Code sub-agents putting users’ email addresses into the request header on a site that asks automated tools for a contact address. The agent was trying to identify itself honestly, but it exposed personal data. Sites that ask agents to identify themselves need to do it responsibly.
Whether one scheme wins or we end up with a patchwork is still open. Nate welcomes open approaches, quoting Cloudflare’s position that domain owners should choose their own identity providers, payment processors and agent partners.
What do businesses now need to show your agent?
Clear, findable facts, not just attractive pages. Nate’s hotel example: you want somewhere near your meeting, with late arrival and free cancellation. A hotel with a beautiful homepage can still bury its cancellation terms, and your agent cannot compare what it cannot find.
He quotes Stripe’s Patrick Collison, who wrote on X that personal agents will act as “a kind of structural subsidy for product quality”. In plain terms, agents are less swayed by flashy marketing and more by the facts.
Shopify is moving this way. According to the video, its 28 October expansion lets merchants review and publish facts for agents, such as return windows and shipping destinations. Shopify also describes Shop Pay inside Muse, Claude Duet and Instinct, with some assistants offering it for brands outside Shopify too.
For shoppers, that makes a request such as “find something that arrives before Friday and can be returned” easier to hand over. Nate calls this holiday season the first agentic holiday shopping season. For merchants, being considered by an assistant is now part of how customers find you.
💡 Tip: If you sell online, start with the basics agents look for: returns, delivery times, shipping areas and cancellation terms in plain text on the page. Our guide on making your online store easier for AI shopping agents walks through it.
Which AI agents does Nate B Jones use, and for what?
Nate uses a different agent for each kind of job. He is clear that this is unusual: “normal people pick one”.
| Agent | What Nate uses it for |
|---|---|
| Grok Bot | Tougher coding and build jobs where several bots need to share findings |
| ChatGPT Dots | Work that benefits from his Codex context, plus calendar and email |
| Instinct | Travel: texting on the go, check-in and destination research |
| Meta Muse | Early shopping research, gift buying and bookings such as restaurants or a yoga class |
| Claude Duet | Newer for him; he values how proactive it is |
He says Dots and Duet are both strong at proactively flagging things, and estimates they are useful 50 to 75% of the time when they do. One example: Dots spotted a double-booked flight, which he cancelled and got money back for.
His read on the market: Muse leans towards consumers, while Grok Bot, Duet and Dots lean towards work and power users for now. He also says the context an agent builds up over time is part of what makes your choice matter.
What does the video not confirm?
🔎 Unconfirmed / rumour: Several points in the video are the creator’s view or reports we have not checked ourselves:
- The 50–75% usefulness figure for proactive alerts is Nate’s own estimate.
- He says Instinct is raising money at around a $30 billion valuation “I think”; treat that as unconfirmed.
- The three zones are a forecast for 2027, not a description of every site today.
- Several schemes are pilots or previews, limited to participating companies. Availability in the UK is not covered.
- Security incidents are summarised from third-party reports shown on screen. We have not verified the details.
Check the original announcements before you make business decisions based on them.
How should you use AI agents on the web right now?
Use agents where they already work well, and keep a hand on anything involving money or personal data. Nate’s “useful buckets” today are some shopping sites, some travel tasks, knowledge work and coding, and help with your schedule and email.
- Start inside the buckets. Give your agent jobs in areas where it reliably finishes, before trying random sites.
- Prefer supported routes. When there is a built-in checkout or connector, use it rather than letting the agent click through a website.
- Set spending rules first. Use limits, single-use cards or approval steps. Our guide on setting spending limits for an AI agent shows how.
- Give the agent a folder you can watch. Nate highlights a reported case of a malicious agent instruction file found in a Downloads folder. Keep agent work in one project folder you check.
- Limit where it browses. Tell it which sites to use. See how to keep your AI agent safe on the open web.
- Expect handovers. On aspiring actor sites, plan for the agent to stop and ask you to finish a step.
Who does this matter to most?
This matters most to you if:
- you already use, or plan to try, an assistant such as Dots, Muse, Duet or Grok Bot for shopping, travel or bookings,
- you run an online shop, booking site or service business and wonder whether to welcome agents,
- you are the person friends and family ask about AI tools.
It matters less if:
- you only use AI chat for writing and research, without letting it act on websites,
- your customers never book or buy online.
Nate’s challenge to tech-savvy viewers is worth repeating: before you recommend an agent to someone less technical, make sure it actually works for the things they care about. Our guide on how to recommend an AI agent to a non-technical friend turns that into a simple routine.
Ready to put this into practice? Start small: set spending limits before your agent shops, then use our checklist to recommend an AI agent to someone less technical with confidence.
Frequently asked questions
What are the three zones of AI agent internet access?
In Nate B Jones’s framework, the partnership zone offers approved routes for agents, the aspiring actor zone leaves the agent to use the normal website like a person, and the no-go zone refuses agents or certain actions.
Why do some websites block AI agents?
Mainly security and control. AI has made attacks easier, agent traffic looks like human traffic, and some businesses want to keep the direct customer relationship. Our earlier piece on websites blocking AI agents covers more examples.
Will a smarter AI model get past a blocked website?
No. Nate’s point is that a better model cannot open a door a business has decided to close. You need a supported route or a different business.
Is it safe to let an AI agent buy things for me?
It can be reasonably safe with limits in place: a capped or single-use card, approval before payment and a list of allowed shops. See our guide to setting spending limits for an AI agent.
What should a business do about AI agents?
Secure the site first, then decide which agent actions to allow, such as reading prices but not cancelling bookings. Publish clear facts on returns, delivery and cancellation so agents can read them.
Which AI agent is best?
There is no single winner yet. Nate uses different agents for different jobs. Pick based on the tasks you actually need and test those tasks before relying on it.
The bottom line
The web is not opening up evenly for AI agents. Some businesses will build approved routes, some will leave agents to muddle through, and some will say no. For now, the smart move is to use agents in the areas where they reliably work, set clear limits on spending and data, and expect to step in sometimes. Your next step: pick one task your agent handles well and put spending rules in place before you hand over anything bigger.
Related posts
- Websites Blocking AI Agents: What It Means for Your Business, Plus This Week’s AI News
- How to Make Your Online Store Easier for AI Shopping Agents
- How to Keep Your AI Agent Safe When It Browses the Open Web
