Day 28 · Bonus 11 May 14, 2026

Generative UI & Agent-Native Interfaces

When the interface generates itself. For thirty years software shipped UI hand-crafted by humans for an unknown user; in 2026 the agent assembles a fresh interface for this user, this task, this moment. Four protocols are racing to standardize the wire between agent and pixel — MCP Apps, AG-UI, A2UI, and the Apps SDK — and the framework you bet on determines whether your product becomes a chat tab or a generative surface.

Read: ~22 min Level: Senior engineer accuracy Day 28 · Bonus 11
Curriculum progress
28 / 28
01 — Core Concept

The interface is no longer a static surface — it is an agent output

For thirty years software shipped a frozen UI: a human designer drew it months before any specific user touched it, the engineer encoded every button in advance, and the application's job was to translate user actions into pre-existing screens. The agent era breaks that assumption. The agent has just read your message, knows what it just retrieved, and can synthesize an interface that fits this exact moment — a chart when you asked about revenue, a form when it needs three more fields, a map when the answer is geographic. The pixel layer becomes a model output.

Analogy first. A traditional app is a printed menu — every dish typeset before the restaurant opens. Generative UI is a chef who walks out, listens to what you want, and plates a custom dish at your table. The menu doesn't disappear (a custom dish for "rice please" is wasted effort), but the chef option exists for everything the menu didn't anticipate. Today's agent products are figuring out which orders deserve the menu and which deserve the chef.

Now precise. Two distinct things hide under the phrase "generative UI" — collapsing them is the most common mistake in the discourse.

1Generative UI (run-time) vs UI Generation (design-time)

UI Generation means an LLM produces React/HTML code that a human reviews, edits, and ships — v0, Bolt, Lovable, Claude Code in design mode. The artefact is source code; the user is a developer; the loop closes in hours. Generative UI means an agent emits an interface description during a conversation that the client immediately renders for the end user — Claude Artifacts, ChatGPT Apps, an MCP Apps response, a CopilotKit useCoAgent output. The artefact is a live widget; the user is an end customer; the loop closes in seconds. The first is a faster way to author traditional UIs. The second is a new product surface.

This distinction is load-bearing in board meetings. "We use generative UI" can mean "our designers use v0" (cost saver) or "our product surfaces dynamic agent-built UIs to customers" (product redesign). Ask which.

2Three implementation modes — code, declarative spec, and component invocation

How does the agent's intent become pixels? Three patterns dominate, each with different safety/flexibility trade-offs:

(a) Code emission — the model writes raw HTML/React/Svelte that the client sandboxes and renders. Maximum flexibility, maximum risk surface (XSS, exfiltration, prompt injection turning into stored XSS). This is Claude Artifacts and the v0 preview mode.

(b) Declarative UI specs — the model emits a structured JSON tree against a known schema, the client renders it using a vetted component library. No arbitrary code crosses the wire. This is Google's A2UI, the original mcp-ui idea, and Microsoft's Adaptive Cards (1.6+ for AI). Lower flexibility, dramatically lower attack surface.

(c) Tool-call to component invocation — the model calls a typed tool (show_revenue_chart(month, segment)), and the application renders a hard-coded React component bound to that tool. The model never touches presentation. This is Vercel AI SDK's "generative user interfaces" pattern, OpenAI Apps SDK's component registration, and AG-UI's frontend tool calls. Lowest flexibility, near-zero risk, easiest to ship.

Mode (a) is the most demoable; mode (c) is what actually ships in production. Most enterprise products will land at (c) for the next 18 months. Treat any vendor showing only mode (a) demos with skepticism about their security review.

3The protocol stack — why GenUI sits next to MCP and A2A

2025–2026 produced a three-layer protocol stack for agentic software. MCP (Anthropic, Nov 2024) standardized agent-to-tool — how an agent calls services. A2A (Google, Apr 2025) standardized agent-to-agent — how agents discover and delegate to each other. The third layer — agent-to-user — has been the most contested. Three contenders ship today: AG-UI from CopilotKit (in-app embedding, frontend tool calls, shared state); A2UI from Google (declarative cross-platform UI specs, v0.9 released May 2026); and MCP Apps / SEP-1865 (an official MCP extension co-developed by OpenAI and Anthropic, formalized Jan 2026, lets MCP servers return interactive UI resources rendered in sandboxed iframes). They are not strictly competitive — many shipping products combine A2UI for cross-surface render with AG-UI for in-app state — but you should know which one ships in the runtime you depend on.

The reference architecture — one diagram

GENERATIVE UI REQUEST CYCLE USER TURN CLIENT RENDERER │ ▲ │ "show me Q1 revenue by segment" │ widget ▼ │ patches ┌────────────────────────────────────────────────────��─┐ │ AGENT LOOP │ │ ┌──────────────────────────────────────────────┐ │ │ │ reason → tool_call → observe → reason … │ │ │ └──────────────────┬───────────────────────────┘ │ │ │ │ │ ▼ │ │ ┌──────────────────────────────┐ │ │ │ UI RESPONSE CHANNEL │ ← pick one │ │ └───────────────┬──────────────┘ │ │ ┌─────────┼──────────────────────┐ │ │ ▼ ▼ ▼ │ │ code blob declarative tool→component │ │ (Artifacts) (A2UI / MCP Apps) (AI SDK, │ │ iframe sandboxed renderer AG-UI) │ │ sandbox typed components frontend │ └────────────────────────────────────────────────────��──┘ │ ▼ RENDERED WIDGET │ ┌───────────┴───────────┐ │ chat-side panel │ │ in-app inline │ │ full-page surface │ └───────────────────────┘ State flow: agent state ⇄ widget state ⇄ user input · AG-UI: shared state primitives (CoAgent state) · A2UI: event handlers route back to agent · MCP Apps: postMessage over MCP transport

4The three axes that separate generative UIs from static UIs

Whenever you compare a generative-UI product to a traditional SaaS product, force the comparison along these three axes — they cut through almost every marketing claim.

1
CARDINALITY
How many distinct UIs does the product produce?
"Three pages designed by humans" vs "a fresh widget per turn"
Static SaaS ships O(10) pages. A generative product ships O(N × turns) widgets — one per task, per user, per moment. Your QA strategy, observability, and analytics must scale to a long tail of one-off views, not a finite page count.
2
AUTHORSHIP
Who decides what appears on screen?
"Product designer at design time" vs "agent at run time"
The locus of taste moves from the design team to the system prompt, the component library, and the policy that decides when to render. Designers in 2026 increasingly ship constraints (allowed components, brand tokens, layout rules) rather than artboards.
3
TRUST SURFACE
What does the user have to trust to interact safely?
"Compiled binary" vs "agent + sandbox + component allowlist"
In traditional UI the user trusts the engineering team that shipped the binary. In generative UI they trust three new things: the agent's intent, the renderer's sandbox, and the allowlist of components the renderer will instantiate. A single mode-(a) raw-HTML escape can re-expose stored-XSS, data-exfil iframes, and clickjacking — vulnerabilities the web spent fifteen years patching out.

02 — Papers to Know

The literature on generative UI is young — three papers anchor the field

Unlike RAG or planning, generative UI does not yet have a single canonical paper. The most useful reading list combines a recent Google Research demonstration paper (proving LLMs can produce production-quality custom UIs), a Working-Definition position paper (the cleanest framing of what GenUI actually is), and a multi-agent UI-generation paper from late 2024 that is the most-cited engineering-flavored treatment in the agentic-UI space.

2026 · Recent arXiv preprint
Generative UI: LLMs are Effective UI Generators
Yaniv Leviathan, Dani Valevski et al. (Google Research) · arXiv:2604.09577 · 2026
First serious empirical demonstration that a modern LLM, given the right tools and prompting, can robustly emit custom interactive UIs that humans prefer over plain markdown output. The system uses Gemini with image-generation and search endpoints, a carefully crafted system prompt with planning guidelines and exemplars, and a post-processing layer that fixes common rendering issues. The team also releases PAGEN, a new dataset of expert-crafted reference UIs for evaluation.
Why it matters. Before this paper, "the model just emits a UI" was hand-waving in talks. This paper turned it into a measurable, comparable, reproducible system — and showed that human raters preferred the model's UI in the vast majority of head-to-head comparisons (still trailing expert-human UIs but comparable on 50% of cases). It is the closest thing the field has to a "RT-2 moment" for UI.
arXiv: 2604.09577
2025 · DIS '25 Conference
Towards a Working Definition of Designing Generative User Interfaces
Kyungho Lee (Expressive Computing Lab, UNIST) · arXiv:2505.15049 · DIS '25 Companion
A definitional paper grounded in a 127-publication systematic review, 18 expert interviews, and 12 case studies. It identifies five constitutive properties that distinguish generative UI from traditional UI design: computational co-creation, expanded design-space exploration, representation fluidity, contextual adaptation, and synthesis over selection.
Why it matters. When you are arguing with a product designer about whether "generative UI" is just "AI-assisted Figma" or something architecturally different, this paper gives you the vocabulary. The five properties are the cleanest checklist for evaluating whether a product is actually generative or is dressing up old patterns.
arXiv: 2505.15049
2024 · Updated 2025 arXiv preprint
Towards Human-AI Synergy in UI Design: Enhancing Multi-Agent Based UI Generation with Intent Clarification and Alignment
Mingyue Yuan, Jieshan Chen, Yongquan Hu et al. · arXiv:2412.20071 · Dec 2024 (v2 June 2025)
Introduces PrototypeAgent, a multi-agent system where a central "theme design agent" clarifies implicit user intent via prompt augmentation, then coordinates specialized sub-agents (layout, component, content) to compose a UI. Targets the iterative-refinement gap in end-to-end UI generation — most prior systems generate once and stop; PrototypeAgent supports human-in-the-loop refinement of intermediate results.
Why it matters. The intent-clarification idea is the single most transferable engineering insight in agentic UI today. Even if you never adopt PrototypeAgent specifically, the pattern of "agent surfaces a clarifying question before committing to a UI" shows up in shipping products (CopilotKit's CoAgents, Claude's "Should I build a chart or a table?" prompts) for exactly the reasons this paper articulates.
arXiv: 2412.20071

03 — GitHub Pulse

The frameworks racing for the agent-to-user layer

The generative-UI ecosystem has six visible centres of gravity in May 2026. None has won — adoption is fragmenting along framework loyalty and target surface (web, native, ChatGPT in-tab).

vercel/ai ~24K
The AI SDK — TypeScript toolkit, unified provider API, streaming, multi-step tool loops, generative UI primitives.
The default starter for Next.js teams. AI SDK 5's "generative user interfaces" doc is the canonical reference for the tool-call → React-component pattern. Distribution moat: ships with v0 templates and most Next.js AI tutorials default to it.
CopilotKit/CopilotKit ~40K
Frontend stack for agents and generative UI — React / Angular, CoAgents, frontend tool calls, shared state. Maker of the AG-UI Protocol.
Raised $27M in 2026 to make AG-UI the standard for in-app agent UX. AG-UI is the only one of the three protocols (AG-UI / A2UI / MCP Apps) with shipping enterprise adoption across LangChain, Mastra, PydanticAI, Agno, AG2 — and cloud distribution via Google, Microsoft, AWS, Oracle.
google/A2UI ~15K
Declarative agent-to-UI protocol — JSON spec, framework-agnostic, web / mobile / desktop renderers. v0.9 released May 2026.
The bet on declarative-spec-over-code. Renderers maintain a catalog of trusted components; the agent can only request renders of allowlisted types. Safer than mode-(a) code emission, more portable than tool-call binding. Co-developed with CopilotKit and an active OSS community.
vercel/ai-chatbot ~19K
Full-featured, hackable Next.js AI chatbot template — streaming, multi-modal, generative UI components built in.
The reference implementation most teams fork on day one. Read it for the simplest production-grade demonstration of how a tool call resolves to a streamed React component in chat.
assistant-ui/assistant-ui ~6.9K
TypeScript/React primitives for AI chat — Thread, Message, Composer, ThreadList — designed to drop into existing apps without re-skinning your design system.
The "headless UI" answer to chat — composable primitives instead of an opinionated theme. Pairs with Vercel AI SDK or any custom runtime (LangGraph, OpenAI Assistants). The right pick if your design system is already strong and you do not want to import a chrome.
MCP-UI-Org/mcp-ui actively maintained
UI over MCP — the origin of the interactive-UI-via-MCP idea. Influenced the official MCP Apps SEP-1865 spec adopted by OpenAI and Anthropic.
Worth reading the source even if you ship on MCP Apps — the design choices around iframe sandboxing, postMessage transport, and resource lifecycle are the practical reference. Shopify's MCP UI post articulated the "breaking the text wall" framing the rest of the industry adopted.

04 — Community Pulse

What practitioners and KOLs are actually saying

Andrej Karpathy · Sequoia Ascent 2026 talk
Karpathy has argued throughout 2026 that the UI layer is shifting toward agents acting on the user's behalf — as a personal example he replaced six smart-home apps with a single WhatsApp conversation to "Dobby," an agent that calls multiple APIs in sequence and reasons about results.
Implication: the "interface" the user touches is increasingly the agent itself; downstream UIs are emitted on demand to confirm or display rather than to navigate. Source: Sequoia Ascent 2026 (karpathy.bearblog.dev), reiterated across his late-2025 / early-2026 posts.
Greg Brockman (President, OpenAI) · X, Apr 2026
"Two big themes of AI in 2026 will be enterprise agent adoption and scientific acceleration."
Verbatim post on X. Context for generative UI: Brockman has framed the human role as shifting from doing work to supervising fleets of agents — generative UI is the supervision surface. Source: x.com/gdb/status/2006584251521839141.
Harrison Chase (CEO, LangChain) · LangChain blog & podcasts, Q2 2026
Chase has argued that agent UX is the next frontier of the agent stack — specifically pushing transparent action logs, the ability to rewind and edit agent decisions mid-execution, and collaborative interfaces that let multiple humans steer one agent. LangChain's agent-chat-ui and Assistant-UI partnerships are the practical expression of that thesis.
Source: LangChain blog tag /harrison-chase, May 2026 "Demystifying Deep Agents" interview series.
Simon Willison · simonwillison.net, Feb 2026
Willison's focus in early 2026 has been on what he calls "agentic engineering patterns" — the working practices for building software with coding agents. He has been less vocal specifically on generative UI, but his "tools that build interactive HTML on demand" posts (around Claude Artifacts) remain among the most-linked introductions to the run-time GenUI idea.
Source: simonwillison.net/2026/Feb/23/agentic-engineering-patterns.

05 — Platform Deep-Dive

How the four major platforms ship generative UI today

None of these four implementations is the same. The architectural choice each made signals where the vendor thinks the future of agent-to-user lives — chat surface, in-app embed, or full agent-emitted application.

Claude
Anthropic
Three-tier generative surface: Artifacts (May 2024) for code, docs, SVG, Mermaid, React; Live Artifacts (April 2026) for dashboards and trackers that re-fetch connector data on each open; On-demand Generative UI (early 2026) that builds working applications — forms, calculators, mini-apps — directly inside the conversation. MCP-integrated, so artifacts can pull from Gmail, Calendar, Slack, GitHub. The Artifacts renderer was open-sourced in May 2026, making it the only major-platform GenUI runtime any team can self-host.
ChatGPT
OpenAI · Apps SDK
Apps SDK announced Oct 2025 and rolled out to ChatGPT Business / Enterprise / Edu in preview Nov 13 2025; developer submissions opened later in 2025, first approved apps shipped to the broader user base in early 2026. The SDK extends MCP — developers register an MCP server plus a UI component bundle. Launch partners: Zillow, Canva, Coursera, Spotify. Distribution edge: reach over 800M ChatGPT users.
Gemini / v0
Google · Vercel
Google's bet is A2UI (v0.9, May 2026) — a declarative cross-platform spec, Apache-licensed, with Workspace add-on tooling for Google Chat. Separately, Vercel's v0 evolved through 2025–2026 from a UI builder into a full-stack app agent — by Feb 2026 it shipped Git, a VS Code-style editor, database connectivity, and agentic workflows. v0 claimed 6M+ developers by Mar 2026. Vercel's framing: "2026 is the year of agents" and v0 is becoming the build-end-to-end-agents surface, while AI SDK 5 ships the generative-UI primitives developers use to expose those agents to end users.
OpenClaw
Skills + Apps
OpenClaw's GenUI story rides on Claude's open-sourced Artifacts runtime plus the SKILL.md hub-and-spoke pattern (Day 14). Skills declare both behaviour and rendering primitives in a single Markdown file; Apps SDK-compatible MCP servers ship interactive UI bundles. The architectural bet is identical to OpenAI's — UI rides on MCP — but with skills as the lifecycle wrapper. The known risk surface (9 CVEs, the 1184 malicious-skill audit from Day 14) extends directly to GenUI: a malicious skill can now emit a phishing UI inside an otherwise-trusted Claude conversation.

06 — Vocabulary

Ten terms to use precisely

Generative UI (GenUI)生成式 UI
A run-time interface synthesized by an agent during a conversation in response to the current task — emitted as code, declarative spec, or component invocation.
Trap: do not confuse with "UI generation" (v0, Lovable) which produces source code at design time.
"Their checkout is generative UI — the agent picks a calendar widget or a date dropdown based on inventory availability."
Declarative UI spec 声明式 UI 规范
A structured (JSON) description of an interface that the client renders using a known component library — no executable code crosses the agent-client boundary. A2UI, Adaptive Cards, original mcp-ui.
Trap: a "JSON UI" that includes raw HTML strings is not declarative in this sense — once raw markup is allowed, the attack surface collapses to mode (a).
"Switch from emitting JSX to an A2UI spec — we cannot let the model write raw HTML in a regulated context."
Component invocation (tool→component)组件调用
The pattern where the model calls a typed tool and the app renders a hard-coded React/Vue/Svelte component bound to that tool. Vercel AI SDK's generative-UI pattern, OpenAI Apps SDK component registration, AG-UI frontend tool calls.
Trap: people call this "generative UI" but the component itself is not generated — only the choice of which component to invoke and with what props.
"We restrict the agent to component invocation — no code emission — for SOC 2 reasons."
AG-UI Protocol 智能体-用户界面协议
CopilotKit's open protocol for connecting agents to humans inside applications — streaming chat, frontend tool calls, shared state, human-in-the-loop primitives. Adopted by LangChain, Mastra, PydanticAI, Agno, AG2 and surfaced by Google / Microsoft / AWS / Oracle.
Trap: AG-UI is about in-app embedding; it does not replace A2UI for cross-surface rendering or MCP Apps for chat-side UI.
"AG-UI gives us shared state between the agent and our React app — that is what enables 'agent edits the form while the user types'."
A2UI (Agent-to-UI)代理到界面协议
Google's framework-agnostic declarative protocol — agents emit a JSON UI tree against a catalog of pre-approved components; renderers exist for web, mobile, and desktop. Apache 2.0, v0.9 released May 2026.
Trap: A2UI is intentionally not Turing-complete — that is its security property, not a limitation to "fix."
"We ship one A2UI spec and the same agent response renders natively on iOS, Android, and web."
MCP Apps (SEP-1865)MCP 应用
The official MCP extension (Jan 2026) letting tools return interactive UI resources rendered in a sandboxed iframe; UI events route back over MCP. Co-developed by mcp-ui creators with OpenAI and Anthropic.
Trap: MCP Apps lets a tool ship UI, but it does not let arbitrary HTML escape the iframe — sandbox + CSP are required, not optional.
"Our Postgres MCP server now ships an MCP-Apps query builder — the user clicks Run inside Claude instead of pasting SQL."
Artifact (Claude) 工件
A side-panel rendering surface for non-conversational output — code, SVG, Mermaid, HTML, React, .docx / .pptx / .xlsx, PDFs. Live Artifacts re-fetch connector data on each open. Runtime open-sourced May 2026.
Trap: "Artifact" outside Claude usually means a build artefact (a binary); inside Anthropic's product it specifically means this rendered output panel.
"Move the weekly report into a Live Artifact so the numbers refresh from Salesforce when she opens it Monday."
Component allowlist 组件白名单
The set of UI primitives the renderer is willing to instantiate from an agent-supplied spec — Card, Button, TextField, etc. The agent cannot request a render of anything outside this set.
Trap: an allowlist of "components" that includes a generic HTMLBlock defeats the purpose.
"Their A2UI allowlist is 23 components — narrow enough to audit, wide enough to express any commercial form."
Shared state (CoAgent state)共享状态
A live, two-way data plane between agent and UI — the agent writes to it, the user's interaction updates it, the UI re-renders. The defining primitive of AG-UI versus pure tool-call binding.
Trap: shared state with no schema is a debugging nightmare — define the contract in TypeScript / JSON Schema before shipping.
"In the AG-UI version the agent watches the user's edits to the draft and adjusts in place — no re-prompt needed."
Agentic surface 智能体载体
A first-class place where an agent can render UI — chat-side panel, in-app embed, full-page replacement, native widget, voice + display. A surface choice constrains protocol choice and vice-versa.
Trap: not every product needs a new surface. For some workflows the chat reply is the surface — adding GenUI adds friction without value.
"We picked the in-app embed surface — AG-UI fits, MCP Apps does not, A2UI is overkill."

07 — Expert Questions

Five questions to drop in a top-AI-engineer conversation

1
If you are building a regulated B2B product today, when does mode (a) — agent-emits-code — ever clear the security review?
Forces a serious answer about CSP, iframe sandbox, postMessage policy, prompt-injection-to-stored-XSS chains. Most defensible answer in 2026: "never in production; mode (c) only, with mode (b) for cross-surface."
2
AG-UI, A2UI, and MCP Apps overlap. Where does each lose and why is none winning outright?
Tests whether they understand the three are not strictly competitive — AG-UI for in-app shared state, A2UI for cross-platform declarative spec, MCP Apps for chat-side embeds. A weak answer treats them as a horse race.
3
How do you evaluate a generative-UI product? "User preference" against what reference, and how do you avoid optimizing for novelty instead of utility?
References the PAGEN dataset / Leviathan-Valevski methodology and forces the hard question of long-horizon engagement vs first-impression preference. The honest answer is that the field does not yet have a great eval — that itself is the signal of expertise.
4
When does generative UI degrade UX rather than improve it?
Senior signal — recognises that variability is a cost. A static UI users have learned beats a slightly-better-on-average dynamic UI for repeat workflows. Honest GenUI advocates say "use it for the long tail, not the head."
5
If Claude's Artifacts runtime is open source and OpenAI's Apps SDK reaches 800M users, what does that imply for the future of standalone SaaS UIs?
The CGO-level question. Forces a position on whether SaaS becomes a backend (an MCP server) with the UI living inside whichever assistant the user already opens. The bullish view: every B2B SaaS becomes a wedge into ChatGPT / Claude. The bearish view: brand erosion and a price war on the backend.

08 — CGO Lens

What generative UI means for product, sales, and competitive positioning at botlearn.ai

The strategic question is not "should we add a chatbot." It is: where does our product surface live in two years — inside our app, inside ChatGPT, inside Claude, or all three? Each answer has different revenue, defensibility, and capex implications.

800M+
ChatGPT users reachable via Apps SDK
3
protocols racing for the agent-to-user layer (AG-UI, A2UI, MCP Apps)
6M+
developers on v0 (Mar 2026) — the dominant generative-UI authoring tool

1Product — pick one surface to dominate first

Spreading thinly across all three surfaces (in-app, ChatGPT, Claude) is a classic Year-1 mistake. For botlearn.ai the highest-leverage first surface is whichever one our power users already keep open. If most of our learners spend their day in Claude or ChatGPT, ship there first using MCP Apps / Apps SDK — the agent does the navigation, we ship the lesson UI as a component. If they spend their day in a learning portal we already own, ship AG-UI inside our app — the agent edits, annotates, and tutors in-place. Pick the surface and resist the rest for 6 months.

2Sales — the demo flip

The sales demo of 2024 was "look at our beautiful dashboard." The demo of 2026 is "ask it anything and watch the UI assemble." A generative-UI demo is dramatically harder to fake — the moment a buyer asks an unscripted question, a static demo collapses. This is a structural advantage for products that have actually shipped the loop; treat the live-demo capability itself as a competitive moat for the next 12–18 months.

3Competitive positioning — the SaaS-becomes-server trap

If our value lives entirely in our React UI, an Apps SDK competitor that wraps the same data into ChatGPT can disintermediate us in one product cycle. The defensive move is to make our data + workflow logic the asset, then ship our own MCP server + component bundle. Customers can still use us through our app, but if they want to live inside their assistant we are there too — and we keep the user account, the audit log, and the billing surface. Treat the MCP server + GenUI components as table stakes for the 2027 enterprise renewal cycle, not a moonshot project.

4Hiring & org — what changes

Design hiring shifts from "ship the next screen" to "ship the constraint system" — design tokens, allowlisted components, layout policies, and example libraries the agent will compose from. Frontend hiring rebalances toward TypeScript/protocol depth (AG-UI, MCP, A2UI) over visual craft. Product management adds the question "what should the agent render here?" to every spec — this is the new RICE column.

5Risk — the safety surface scales with creativity

Every additional unit of UI freedom is a new attack surface. A model that can emit raw HTML can be coerced (via prompt injection in a learner's pasted document) into rendering a phishing form inside our own brand. Day 11's threat model now extends into the UI layer. Concrete asks: ban mode (a) in production, mandate component allowlists, treat any tool that returns text destined for an LLM plus a UI render as a privileged operation, and red-team the GenUI surface explicitly before any external rollout.


09 — Curriculum Tracker

28 days · the agent stack from theory to interface

D01Full Agent Stack
D02Memory Architecture
D03Planning & Tool Use
D04RAG Deep Dive
D05Agent Frameworks
D06Benchmarks & Eval
D07Multi-Agent Systems
D08Computer Use Agents
D09Code Agents
D10Long-Horizon Tasks
D11Agent Safety
D12Agent Economics
D13Research Frontiers
D14OpenClaw Deep-Dive
D15A2A Protocols
D16Agentic Commerce
D17Synthesis
D18Reasoning Models & TTC
D19Agent Observability
D20Voice & Realtime Agents
D21Agent RL & Fine-tuning
D22Context Engineering
D23Open-Weight Agent Models
D24Deep Research Agents
D25Async Agents
D26Agentic Browsers
D27Embodied & VLA Models
D28Generative UI & Agent-Native Interfaces
Day 28 closes the surface trilogy started on Day 8 (computer use), continued on Day 26 (agentic browsers), and capped on Day 27 (physical embodiment). Tomorrow's possible directions: world models for planning, agent identity / authn infrastructure, or vertical-AI go-to-market patterns. Pick the one your fund-raise narrative needs first.