--- title: Agents at the Gates date: 2026-10-13T08:46:27Z modified: 2026-10-13T15:50:04Z permalink: "https://dgw.ltd/2026/10/13/agents-at-the-gates/" type: post status: publish excerpt: "" wpid: 1854 categories: - Agents - AI featured_image: "https://dgw.ltd/wp-content/uploads/2026/10/agents-at-the-gates-og-1.png" featured_image_alt: request → confirm → book Agents at the gates timestamp: 2026-10-13T15:50:04Z tags: - Agents - AI --- A friend sent me [this NY Mag article](https://nymag.com/intelligencer/article/ai-agent-flood-muse.html) recently, titled _Nobody Is Ready for the AI-Agent Flood_. It explains a near (and emerging) future where AI agents are using websites and services on behalf of their human users. A few things from the article stood out: Businesses began getting calls—actual voice calls—in September from Muse, Meta’s assistant, released only last month. The assistants opened calls with “Hi, this is \[Name\]”, said they were calling on someone’s behalf, and disclosed “I’ll be transcribing this call”. Other requests included a quote to install three bidets, a karaoke room and “a cake order for pickup today”. A Seattle karaoke venue employee asked “Are you AI?” and said “I need the real person.” Later, she was quoted as saying to the journalist: “We don’t like AI calling our restaurant. We hate it.” Researchers looked at 84 suspected AI-driven spikes in the use of online government services and named the pattern **agentic flooding**: what happens “when frictions that were de facto rate limits disappear”. AI assistants routinely phone businesses, fill in forms and make bookings for people, and will be doing so in larger and larger numbers, as the above hints at. For the person, it saves effort. For the business on the other end, it removes the thing that used to keep demand manageable. tl;dr: asking took effort. > “My biggest worry,” said Schmitz, “is that if governments really want to, building in friction is cheap, has precedent, and will be very attractive on short notice.” We can try to detect and block assistants, but often this is a losing game. Many agents don’t [identify themselves](https://labs.chancerylaneproject.org/2026/10/07/the-hits-keep-coming/) anyway, and if they do and get knocked back, they’ll likely spin up a browser and pretend to be a real user anyway, or use [curl](https://curl.se), or find the service’s API and create (or [delete](https://thenextweb.com/news/openclaw-ai-agent-gym-booking-api-flaw-australia)!) an entry that way. Further, attempts to block agents could potentially reject or frustrate real customers when they guess wrong (increasingly complex and broken CAPTCHAs already point to this arms race). I was intrigued/terrified by all of this, as well as recognising a pattern. Along with the Chancery Lane Project, we have been building a tool, [Markdown for Agents and Statistics](https://wordpress.org/plugins/markdown-for-agents-and-statistics/), and so have seen first-hand the actual numbers that we think are agents. And that’s the rub: loads of what we think are bots are “unknown”. We just don’t know what a % of the visitors to the website actually are. Tom Loosemore also [found this issue](https://loosemore.com/2026/10/09/how-agents-show-up-what-muse-and-other-agents-reveal-when-they-use-whenisbins/). Meta’s Muse visited a service he had built, [WhenIsBins](https://loosemore.com/2026/10/09/how-agents-show-up-what-muse-and-other-agents-reveal-when-they-use-whenisbins/). He noted: _No header, token or signed request says “Muse”_ He’s also written extensively about how important it is for agents to identify themselves, so we as service designers can build services for an increasingly large share of website visitors, namely agents. Increasingly, agents, whether Muse, Dots or Grok Bot (urgh), will be visiting sites. If they can’t get in, they will find another method. ![Markdown for Agents and Statistics dashboard showing agent visits by number Total agent visits: 1,217 On-demand reads: 369 Search: 197 Training crawls: 46 Unknown: 605 ](https://dgw.ltd/wp-content/uploads/2026/10/RTzX4OKzJsmX5nRu9emrc1zjhNMEEohiq1u56ZO7-1024x614.jpeg)Markdown for Agents and Statistics dashboard showing the number and type of agent visitsThe earlier article also talked of the pressures this will put on businesses, whether through automated calls (via an AI-generated voice), booking forms being populated or even, in one case, managing to convince a brewery to open 2 hours earlier than usual to accommodate someone’s AI-managed itinerary. The risk of no-shows goes up exponentially when the user isn’t even the one organising it. It sounds quite weird, the idea of outsourcing the booking of a restaurant table to AI. But people are doing it. And if the aforementioned AI agents with their mascots and user-friendly interfaces take off, the agents are coming. Small businesses in particular are very susceptible to this new type of visitor. A thing I’ve been thinking about for a while now: what does UX design look like for a user that is not “using” the website? They “see” it in a sense. But their goals are different. Further, the business’s goal is also different and potentially contradictory. They want users to successfully complete a goal: book a table, order a product, file a query. But they also want that to be genuine. If an agent is acting on behalf of a user, this may not be the intended outcome, and it might not be flagged until it triggers a cancellation, a no-show, a return or an unnecessary response. So what sort of UX patterns might we consider for the age of the agents? One pattern is one I have already delivered. [Markdown for Agents and Statistics](https://wordpress.org/plugins/markdown-for-agents-and-statistics/) serves optimised content to agents and bots in the form of Markdown. By stripping out the images, styling and other page content not considered necessary for a bot, we deliver content that is roughly 80-90% smaller in size. This, we hope, leads to less energy use and greater efficiency for the AI agent reading the site. It also means our servers are not getting hammered by an AI agent. So route one: optimise your content, serve [Markdown instead of HTML](https://developers.cloudflare.com/fundamentals/reference/markdown-for-agents/) (Cloudflare offers this, as does the WordPress plugin we have built), record what and who visits your service. So what else? Here are five more ideas as clickable demos, mostly small, but I think we’ll likely see more techniques, non-user journeys (AI journeys? Bot journeys?) and design patterns emerging as services increasingly deal with agents. These are very rough and more intended to provoke discussion. [Confirm with the person](https://dgw.ltd/wp-content/themes/dgwltd-theme/dist/tidewater-confirm.html) An order is held until the customer replies to a text. [A page that answers before the call](https://dgw.ltd/wp-content/themes/dgwltd-theme/dist/tidewater-ordering.html) One page with the facts an assistant rings up to ask, in plain words and as Markdown. [Count what arrives, not only what is read](https://dgw.ltd/wp-content/themes/dgwltd-theme/dist/tidewater-count.html) Count the requests that come through the form, and how many a person confirmed. [Ask for a deposit](https://dgw.ltd/wp-content/themes/dgwltd-theme/dist/tidewater-deposit.html) A small deposit for anything that holds a slot, and names out of a hat in place of a race when slots are scarce. [On the phone: a short script](https://dgw.ltd/wp-content/themes/dgwltd-theme/dist/tidewater-card.html) A card by the phone: take the customer’s own number and confirm by text. These all become semi-permeable barriers to agents without disrupting existing user journeys. It might mean some more friction, for example, a small deposit before a booking is confirmed. A reply to a text or email, say. But it might well be necessary to avoid the problems this new type of user poses and ultimately can help all parties deliver and use services. More than anything, we need to collectively consider these approaches when designing services in an age where we are not 100% sure of the provenance of our users.