interfaces for humans
NewField report No. 1: the plumbing settled, the consent layer didn't

Learn how AI builds interfaces.

AI products now answer with screens, not just text, and agents now use apps built for people. This free field guide explains how it all works, and how to ship it safely.

Start learning →Try a lab

Pick your way in.

  1. ReadThe curriculum8 short modules, from the basics to safety and evaluation. Start at zero; no prior AI work needed.Start module 00 →
  2. PlayInteractive labs9 labs that show what happens between the agent and the screen. Scrub through it step by step.Browse the labs →
  3. Keep upThe field reportOne letter a month: what shipped, what broke, what changed. Dated and sourced, five minutes.Get the newsletter →

The big picture

Seven layers, one map.

Every AI-generated screen passes through the same layers, from the model at the bottom to the person at the top. Once you see which layer each tool lives on, “which one should we use?” mostly answers itself.

  • Read down: how an AI answer becomes a screen.
  • Read up: how an agent reads and uses a screen.
Learn the stack →
Fig. 1 — The generative UI stacked. 2026.11
  1. 06Personattention, intent, consentthe only layer that can say yes
  2. 05Host & sandboxwhere generated UI is allowed to runMCP AppsOpenAI Apps SDKartifactsiframes + CSP
  3. 04UI descriptionwhat to draw: components, specs, or codeA2UIjson-renderOpenUItool → component
  4. 03Agent ↔ UI transportevents, state deltas, tool calls, interruptsAG-UISSE / WebSocketAI SDK UI stream
  5. 02Agent runtimeplans, graphs, memory, checkpointsLangGraphGoogle ADKMastraAI SDKCrewAI
  6. 01Tools & other agentsAPIs, data, delegated workMCPA2A
  7. 00Modeltokens: text, JSON, codestructured outputtool calling
reverse channel · WebMCP · a11y tree · computer use
Read top-down for how an answer becomes a screen; read the blue rail bottom-up for how an agent reads and operates a screen. Most confusion about “which protocol wins” disappears once you see they sit on different layers.

Curriculum

Start at 00. Stop wherever you need to.

Each module is a short essay with its labs and sources. The first three get you to a working mental model; the rest take you to expert.

Start module 00 →

Popular labs

  1. 00What generative UI is (and isn't)8 min
  2. 01The raw materials7 min
  3. 02The protocol stack11 min
  4. 03Where state lives10 min
  5. 04Building it for real10 min
  6. 05Interfaces agents can use7 min
  7. 06Trust, consent and evaluation8 min
  8. 07The frontier7 min

Writing

Latest.

All writing →
  1. field report · 2026-11-10 · 4 minField report No. 1: the plumbing settled, the consent layer didn'tThe first monthly letter from Interfaces for Humans. A2UI hit 1.0, OpenAI's Agent Builder shuts down on November 30, half of model-generated components ship a dark pattern, and one pattern to steal for approvals.
  2. essay · 2026-11-08 · 8 minThe model only proposesThe fake "sent" tick, the payment that fires before the dialog paints, the agent reasoning about a filter you already cleared — they're the same bug. A four-plane map for building generative UI that doesn't lie.
  3. field report · 2026-11-07 · 5 minThe state of generative UI, November 2026Generative UI went from demo to default this year. The protocols consolidated into plumbing, the SDKs churned, and the one layer nobody standardised is the one that decides whether people get hurt. A field report with dates.

Shipping generative UI into a real product? We audit what your generated screens and agents actually do before they reach users.

Work with us →