DDX is a UX-focused design and innovation conference that takes place around the world. Founded by Sebastian Gier from the podcast he bootstrapped in Munich, the conference landed in Miami for the first time last week. Sebastian said this one would test the city. The result: it’s coming back next year. With any luck, so will I.
12:00. Meghan Preiss, Head of Global Service Design, is on a panel about leading design in an agentic world. Her team at a themed entertainment company (name redacted, but its one you all know) has pulled its system map out of FigJam and Miro and onto paper, so people can see the whole system at once and touch it with their hands. “This week I wrote a business case asking for a plotter printer.”
5:30. Gleb Kuznetsov, Chief Design Officer at Brain Technologies, puts up a slide labeled “My career in one slide.” Corner radius, 2010 to 2035. Four pixels. Eight. Sixteen. Twenty-four, captioned “We called it innovation.” Then a dashed shape for 2035. r∞. No corners, because there’s no screen.
A machine that puts the system on paper at human scale.
A screen rounding itself out of existence.
Between them: the day’s obscured subject.
The Human API.
an oxymoron
The day before, Meghan and I were at the University of Miami as part of a guest lecture series co-hosted by DDX. I lectured for Sanne Martens’s class, AI and Creativity. In Beyond the Prompt: A Cyborg Foundation for Interaction Design, the thesis I brought into that room was the same as this publication: we are all cyborgs. One subassembly of that argument was something like this:
The oxymoron is Human API. API is an Application Programming Interface, a surface for two different software programs to talk to each other. Humans use UIs. Programs use APIs. But when we tear down the veil between human systems and technology systems, UI and API become different interface types inside the broader cyborg system. One bleeds into the other. In that way, the Human API is a very real thing.
When the term was raised at the last DDX keynote, my ears perked. In my June systems interview I described augmentation AI as “coding it while it writes back to our human API.” (Please forgive me the excess of acronyms). I was talking about myself.
Back to DDX: a full day of people who design interfaces for a living, or at least play some part in evaluating or bringing them to life, contending with the unfolding realities of working in intelligent machinery. I missed the morning keynotes, Meghan’s and Cassie McDaniel’s, to last-minute work. I came in at the roundtables and stayed through the final slide.
// sys.log operator claims to be an interface. requests a test. schedules it for the next morning. arrives late to it.
You, the API
In Gleb Kuznetsov’s “The Last Screen,” he told a story about waiting two months for eleven minutes with a doctor, forgetting nearly all of it on his delighted walk to the car (nothing serious, it turned out), then attempting to uncover the pieces between a notes app, a prescription, and a pharmacy that never talk. Then:
“You are the API. You are the interface.”
When he said it, he meant it as a diagnosis. The human is the glue, the translator between apps, and he wants to retire the human from that job. Early in the talk he turned to trust, to how quickly people have come to rely on AI. An amazing shift “because before we never trusted technology at all.”
I heard the cyborg condition in his line. The rest of the day complicated both readings. If the human is an API, it still exposes endpoints. What many speakers through the day wrestled with without explicitly stating it is that, while we are still a Human API, we are designing the endpoints, the terminals, that remain.
AI evaluator
In the age of intelligent machines, trust and behavior form the matrix. KPMG and the University of Melbourne surveyed 48,340 people in 47 countries for their 2025 global study. 46% said they were willing to trust AI. Two in three workers (66%) said they had relied on AI output without evaluating it.1
Elizabeth Reme, Senior UX Researcher at American Express, ran a workshop on these axes. She provided me and other workshop participants with twenty-four customer comments from a made-up meal-kit company. We sorted them by hand first, then rated our confidence on a scale from “Guessing” to “Certain.”
Then she showed us what Claude Opus 5 did with the same comments and a one-line prompt. It was fast.
It flagged a sarcastic comment for what it was, something LLMs can struggle with. But it also wrote, in relation to apparent bugs in the app: “Most customers tie these to the recent update.”
But only two of twenty-four did.
She presented a trust-building pipeline: Step three is human review, where a person checks claims, causes, and severity before anything ships. Her tuned prompt, the one behind a second run, tells the model to “Rate severity (safety, privacy, money) as well as count.” The slide footer: “The prompt is step 2. The rest is what I build for teams.”
After the workshop I had a chance to walk and talk with Elizabeth. She described a key to the puzzle, what she calls her “golden standard”: a reference set of her own interpretations of customer data that her research agents learn from. Every run, she learns, and some part of it gets edited. She described the artifact, roughly, as living.
John Maeda put “Evaluation Is All You Need” first in his Design in Tech Report this spring at SXSW.2 His examples came from fields with right answers. Elizabeth writes the provisional right answers herself, in a field that rarely has them.
The evaluator endpoint. A person reads what the machine concluded, and signs. The all-important signature Meghan called out in her talk at UM.
second-hand consent
Sures Kumar, Staff Builder at Google DeepMind, gave an agent “his whole life” for 6 months. Years of email and calendar. Ten years of transactions. Journals, lab results, step counts, even select text message streams. He called it a professional stunt, the kind you’re not supposed to try at home. It took three months before he trusted it enough to walk away from it.
Messages were the careful part. His slide read “Consented individuals only.” Friends and family opted in.
“We should get the consent of the persons in their own way for my agent to access the data.”
Most consumer apps, he noted, skip that step. You click a button called Connectors and hand over everything.
The result of his test was an intricately constructed design for one, chock full of beautiful data visualizations. His sleep became a garden, depth drawn as a rose, a good night as a bloom. “When you design for one, you can really design for one.”
Once a week, he would check how the agent is doing “by opening behind the screen curtains.” You have to choose the tool that shows how it thought. He implied that some of the AI tools we already know do this better than others, and offered it as the test for choosing an agent in the first place. Of the app that makes everything easy and fast instead, he said:
“That is not user experience. That is a way of just hiding the parts that need transparency or auditability.”
The consent endpoint faces two ways. The owner agrees to be known. The people inside the owner’s data need a say too.
// sys.log consent boundary detected. owner is not the only exposed surface.
a rift in the timeline
Monica Girel, Senior Product Designer at Adobe, redrew information architecture. “IA for AI is a timeline.” Traditional IA maps space. Hers maps phases: entry, refinement, execution, output. On her slide for a long-running agentic task, execution is fashioned with a stop button and visible reasoning. Then a coffee break. Then “User leaves and comes back.” Then a reasoning log.
The stop button was the smallest endpoint of the day. Sures had named its absence earlier that afternoon: “there’s no turn off toggles for agents.”
In the Q&A, Daniela Grosz3, Lead Product Designer at Flow, pushed on the frame. The foundation is time-based, she said, “but the model itself seems to be quite linear.” What about experiences in parallel, repeating, skipping ahead, looping back? Monica held to chronology, one user and one experience at one time. Layer a second timeline. Draw two. “I challenge you to make a fork in the road.”
I read the exchange as a working model doing its job. The timeline holds. It also has room to grow. Daniela’s question gets sharper once one person runs several agents at once.
Then the human API has concurrency.
Three timelines.
One pair of eyes.
A stop button on every thread, and one coffee break shared by all of them.
The model needs the same behavior as the work: branching, forking, merging. A timeline in a sandbox, branching, forking and merging just like our code.
// sys.log concurrent sessions on a single human endpoint: 3. queue: none. interrupt handler: shared. the human calls this multitasking.
the builder endpoint
The night before the conference at the University of Miami pre-event, Cassie McDaniel, VP of Product and Design at Medium, surfaced something I see teams around me struggling to embrace.
“All of our designers are now submitting code to our codebase. For better or worse, our engineers are also making product design decisions.”
Monica, on the same panel, said her designers push code to production too.
Then Cassie polled the room. Who had vibe-coded a tool that helped them? maybe half the room eagerly shot their hands up. Who used a tool someone else had vibe-coded? one lone hand.
Her read: the tools are democratized, and discovery is the hard part. Mine, scribbled in the margin, was that half the tools I use were built with AI augmentation and growing rapidly. Both can be true.
At the final conference panel I posed a question to the group: had anyone else stopped making slide decks, or other artifacts? With the absence of some legacy design artifacts, and the availability of rapid tooling, what new stuff are you making? Raul Justiniano, who leads customer design at Asurion, answered: “prototypes, high fidelity prototypes are now the center of the conversation.” A researcher rebuilt their personas as LLM-enlivened avatars you can question in real time. When one has no answer, it pings the researcher.
Cassie, the night before: “I think the hard part has shifted more towards the end for us.”
The builder endpoint. The Human API now writes its own clients. Meghan’s plotter belongs here too. It is an output endpoint, a way to hold the whole system in a room at once.
// sys.log user-generated tooling detected. artifact layer shifting from deck to prototype.
context in the Flesh
The last panel was titled “More Human, Not Less.” The moderator, Jonathan Ruiz of DocuSign, who organizes Friends of Figma in Miami, asked a question and then repeated it for the room. “What is the line between AI making you more yourself, and AI quietly becoming you?”
Raul described hiring rounds where candidates showed polished AI process and couldn’t say what problem they had solved. “AI cannot replace that because you were the context.”
Another panelist said to understand what you enjoy about design, “because you don’t want to get that from AI if you enjoy it.”
Sures had said it after lunch, more bluntly: “please do not automate anything that you enjoy doing.”
Obvious, maybe. Harder in the flurry of dancing with the machine.
Then Cassie, from the audience. “In the olden days, I would build trust with my partners, seeing their design decisions, seeing how they write docs, seeing how they craft messages. I feel like that’s getting lost.”
My notes from that hour say “oddly philosophical and existential.” It felt like a break from the tooling talk. Context is the endpoint that cannot be exported.
// sys.log context transfer failed. local embodiment required.
the last screen
In the Q&A after Gleb’s talk, I took the mic. Sures had teased this idea of designing for one after lunch, then underlined a lack of what I’ll call second-hand consent. Cassie pointed to it the night before when polling the room about using their own vibe-coded stuff versus others’.
I put it to Gleb: “You mentioned designing for one, and their people. That’s the second time that’s come up today.” He answered, roughly: once your mail can talk to your messenger, the system learns you have a brother, a mother, what they like, and it can build interfaces that serve them as well.
His own deck had already drawn the line. The autonomy contract, under “Must ask”: “Before irreversible. Before sensitive. Before affecting others.”
Then he dated the screen’s end and described the one that survives. “The last screen we design will be the only one where it’s going to ask you something very, very, very important and very deep.” Something that “cannot be undone.”
That is the strange part. The man who drew r∞ ended on a screen.
It called back to Meghan’s “Wet Noodle” from the UM lecture before mine: the thoughtless, glazed state scribble on a touchscreen, signing away data and consent without bothering to form a letter, never mind a name.
In June, Maeda wrote that screens become “where judgment happens.”4 Gleb’s 2035 keeps one, and it is a judgment screen. The glue job goes to the machines. The calls that stay are the ones this day kept designing: evaluate, consent, interrupt, build, carry context. The last screen is the final one. A yes or a no, from a person, about something that will not come back.
At SXSW in March I wrote that “the cyborg is the entity that has learned where to sit in the loop.” Miami was a room full of people working out the seating chart.
The human API deserves its own spec.
Inputs.
Rate limits.
Error codes.
The terms of its contract.
// sys.log one endpoint left unmapped in this file. it is the one reading it.
Notes
Assembled by Marcel with Cache, Ink, Grep, Sol, Wire.
John Maeda, Design in Tech Report 2026: From UX to AX, SXSW, March 2026. “Evaluation Is All You Need” is lesson 1. designintech.report. Covered on Signal+Static in SXSW Field Log // Day 7: Human in Which Loop?↩︎
Daniela also has an amazing and growing UX and AI related Insta: insta/@uxdannydesigns - go check her out.
longer, deep dive available and now open to free subscribers: https://signalstatic.substack.com/p/beyond-the-edge-the-full-ddx-nyc