Agents hate bad UX too
Stephen Patterson, a designer at Deck, a company that builds infrastructure for computer-use agents, writes on UX Collective that agents get stuck exactly where people do: ambiguous buttons, menus that only appear on hover, sessions that expire in the middle of a form. In his example, an agent picked the wrong one of two buttons labelled "Continue" and "Submit." People muddle through bad design. Agents can't, which makes them, as he puts it, "the most honest usability testers we've ever had."
https://uxdesign.cc/agents-struggle-with-the-same-bad-ux-we-do-98415f2b8981
This is a very interesting story. UX is just as important in the AI world, perhaps even more important, if customers use them via agents and you have to trust them to navigate a website. The irony here is hard to miss, that UX is becoming even more important now that machines use our websites.
Vercel's agent-desktop, which topped GitHub trending, shows where this goes. It lets AI agents read and operate desktop apps on macOS, Windows and Linux through the operating system's accessibility tree, "without screenshots or pixel matching." So an app with good accessibility labels is also an app an agent can use. Who would have thought that this is why accessibility and UX might be getting more important now?
https://github.com/vercel-labs/agent-desktop
Nolan Lawson had agents build the same Pokémon app in 12 JavaScript frameworks and measured what it cost. Hono with htmx was cheapest at a median of $2.67 per run, Angular the most expensive at $4.97, and Astro the fastest at 8.2 minutes. He warns that results vary a lot from run to run and that it's one small app. Indeed an interesting test, and something I would like to see more of. The language itself is now just a question of how well agents can work with it. Who cares about what code something is written in if nobody will ever look at it?
https://nolanlawson.com/2026/10/28/the-ai-cost-of-javascript-frameworks/
And for briefing your agent the way designers brief people: Scrawl lets you show an agent the problem with a visual frame instead of describing it. I haven't tried it yet; so far there's only the launch page.
https://www.producthunt.com/products/scrawl-2
When the machine does the math
On Terence Tao's blog, mathematician Frank Calegari calls OpenAI's release of more than 700 research papers on October 6 "a watershed in the history of mathematics." His conclusion: "Choosing worthwhile problems may end up mattering more than producing proofs." I like the angle here. This is a development a lot of jobs will face soon.
https://terrytao.wordpress.com/2026/11/01/mathematics-has-changed-lets-act-like-it/
A second essay on the same blog goes further: mathematicians should stop tying their worth to being the ones who make the discoveries. "Our purpose rests on our willingness to let truth govern what we seek and what we become." Replace "proofs" with "pixels," and you have the designer's version of the question.
https://terrytao.wordpress.com/2026/11/02/egoless-mathematics-in-the-ai-era/
Designer Christopher Butler answers it for our field. The fear of someone faster, cheaper and better is old, he writes, AI or not. Clients can get the work faster and cheaper elsewhere. "But they don't have me. And they don't have you. Be singular." Interesting and true.
https://www.chrbutler.com/faster-cheaper-better
Also worth a look
HackerBooks, and what a "recommendation" is
A day after Karpathy's HackerTube, a Hacker News user built HackerBooks: books mentioned by Hacker News commenters, mostly written by Claude. The catch: counting mentions treats criticism as a recommendation, so at one point Mein Kampf had been "recommended" by 651 HN users. Fun project that shows the common problems now. The fast part was the code. The judgment about what a mention means was not.
https://news.ycombinator.com/item?id=50326821
Siri swallows your apps
With the 27.3 betas, any developer can connect their app to Siri AI through App Intents, so Siri can use their content and actions directly. Apple is trying to swallow its whole ecosystem into Siri. It will probably work, because it's Apple. (Not in the EU, of course.)
https://developer.apple.com/news/?id=gdk2umrf
Nexus Mods labels AI instead of banning it
Nexus Mods, the biggest site for PC game mods, will require authors to declare whether they used AI, and let users hide AI-generated mods. It "won't introduce an outright ban on AI-generated mods," because AI is now used throughout the whole process of creation. Interesting approach, but I guess that everything there will soon be labeled AI-assisted, which makes it meaningless again.
https://www.theverge.com/games/1026085/nexus-mods-ai-disclosures
Google's tutor, graded by Google
In a Google trial with more than 300 students at the University of Tokyo, those using only standard web tools improved by 5.3 percentage points in one study session, and those with Gemini's Guided Learning by 8.1, which Google calls "1.5x higher learning gains." Very interesting, but not very surprising.
https://blog.google/products-and-platforms/products/education/guided-learning-uk-japan/
Have you watched an agent fail at an interface you designed? I'd like to hear what it got stuck on. Leave a comment here or reply on Bluesky.