26 September 2026
Heard in AI

Dave Blundin says websites should serve AI agents and still let people check them

On the Moonshots podcast, panelist Dave Blundin argued that banks, healthcare providers and online stores should give AI agents clean, machine-readable interfaces instead of making them scrape pages built for people. He does not expect the ordinary website to disappear. In his view, it will remain as a parallel view where people can check what their agents saw. This is his design proposal, not an adopted web standard. He illustrated it with a memory of Amazon that a firsthand account describes differently.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on Moonshots with Peter Diamandis, episode published 22 September 2026

An audience question on a Moonshots AMA episode came from someone who works on how customers deal with banks and healthcare services online. Host Peter Diamandis passed the floor to Michael, who said he runs a research consultancy focused on digital customer experience in financial services and healthcare. His worry was simple. If everyone soon has a personal AI agent acting for them, the consumer website and the mobile app look to him like they are "getting cooked." So what will dealing with your bank, your doctor or Amazon look like in the near future?

The episode was published on September 22, 2026. Dave Blundin, one of the panelists on the show, answered. He proposed that companies build two front doors: one for agents and one for people.

Agents today read pages meant for human eyes

An AI agent is software that carries out tasks for someone, such as checking a balance, booking an appointment or placing an order. Both Michael and Blundin used the nickname "Skippy" for this kind of personal agent. Today, Blundin said, a lot of what these agents do "happens through screen scraping the website." The agent reads a page designed for people and tries to work out prices, buttons and account details from its layout. He called that approach "slow, but also error prone."

His fix is for every site to also offer a structured, machine-readable interface. He described it in terms of XML, a long-established format that labels data so software can read it reliably. Landing pages, he said, should effectively announce: "if you're an AI or a bot, look here." Behind that signpost would sit "all the data beautifully formatted," so an agent can get what it needs and complete whatever transaction its user wants. In his view, anyone designing a new customer interface now "really ought to be thinking an agent needs to be able to self-serve" from it. He placed this in the "very near future."

Why he does not think the website is finished

Blundin rejected Michael's premise that the website is cooked. His reason has less to do with people browsing and more to do with people checking.

Even when an agent does all the work, he argued, its user will want to ask where a piece of information came from. The agent can point to a web address, and the person can click it and see the source. He called the human-readable site "a parallel view for humans of what the agent can see," one that keeps an agent's work "visible and transparent to the human operators." In his words: "I think it just runs in parallel."

In Blundin's picture, the machine-facing interface handles the transactions and the human-facing page lets people check the agent's work. He did not address how the two views would stay consistent or what an agent should be allowed to do once it gets in. The episode dealt with security risks in a separate discussion.

The Amazon story, and a firsthand account

Blundin supported the idea with a story about Amazon. As he recalled it, Jeff Bezos came back from one of his walkabouts many years ago and demanded an XML interface on every database and system at the company. The goal, as Blundin told it, was to let Bezos personally spot-check any part of the operation through his browser. According to Blundin, the IT staff warned that it would be very slow and everyone revolted, but Bezos forced it through anyway.

Steve Yegge's firsthand account of that period, in his Google Platforms Rant, describes the idea differently. Yegge recalls Bezos issuing a service-interface mandate around 2002, though he is unsure of the exact year. Under that mandate:

  • Teams had to make their functions and data available through interfaces that could be reached over the network.
  • Teams had to communicate through those interfaces instead of reaching directly into other teams' databases.
  • The interfaces had to be designed so they could one day be opened to outside users.

Yegge's account does not require any particular technology. He lists HTTP, CORBA, publish–subscribe messaging and custom protocols as acceptable, and does not single out XML. He also does not describe the mandate as a way for Bezos to inspect systems in a browser. He writes that the resulting architecture required serious investment in monitoring, clear ownership, usage quotas and throttling.

The principle that Yegge documents, and that fits Blundin's argument, is that every system should offer a clean, well-defined interface that others can use without depending on its internal details. The details about XML and browser spot-checks come from Blundin's own memory.

What this would mean for a bank or a clinic

For the kinds of businesses Michael studies, Blundin's proposal comes down to two parallel surfaces. One is a structured interface that a customer's agent can use to find information and carry out transactions without guessing at page layouts. The other is a human-readable site that stays in place so customers can follow the agent's trail and confirm what it relied on. This is Blundin's recommendation, not a standard that banks, healthcare providers or retailers have announced they are adopting. His prediction is that the website will remain, running in parallel as the place where people check their agents' work.

Share this article

Go to the original

Sources & further reading

  1. 01

Connected ideas and articles

From the conversation

Podcast episodes

Moonshots with Peter Diamandis

Ask the Mates Anything Round #2 | MOONSHOTS AMA #293

Episode published This article draws on 54:00–56:20 (approximate times)

Article history

Updates to this article

Tags

Box inspects its heaviest AI users to find workflows worth teaching

Box CEO Aaron Levie says the company keeps a list of who burns the most tokens — not to encourage more spending, but to check whether the usage is waste or a practice worth demonstrating to everyone else. He describes pulling a team into a room within six hours to watch one colleague work, reports two-to-threefold gains in delivered customer-facing functionality in parts of the stack, and explains why Box will not drop code review.

6 min read

What changes when an AI agent gets its own computer

On the Moonshots panel, Peter Diamandis runs a Grok Bot chief of staff called Skippy and Emad Mostaque runs 18 of them across his own machines, installing models and making art. Salim Ismail calls it the move from asking an AI to assigning work; Alex argues the messaging-app interface cannot possibly scale.

6 min read

Robinhood's AI trading agents mostly remove the legwork for now, Vlad Tenev says

On the Moonshots podcast, Robinhood CEO Vlad Tenev said the company's Agentic Trading accounts are currently used largely for automation: agents assemble and execute complex trades that customers have already decided on. He did not claim they give users an edge over professional quant funds. He said general-purpose AI models are often reluctant to trade and that matching a sophisticated quant fund's infrastructure is a long roadmap. Robinhood's own documentation adds that connected agents can read data across a customer's accounts and, if authorized, trade without confirming each order.

8 min read

ChatGPT gets the patient chart, and a panel wants AI double-checking diagnoses

OpenAI's September 1 announcement lets healthcare organizations connect authorized Epic patient records to ChatGPT, so clinicians can ask what changed since a visit, review labs and medications, and find referrals that were never closed, with summaries pointing back to the chart. On Moonshots with Peter Diamandis, Emad Mostaque called for a sprint to have every health decision double-checked by an AI within a year or two, and another panelist predicted it would become malpractice to diagnose without AI in the loop — proposals, not current clinical practice.

5 min read