--- title: "What Is Agent-Native?" description: "Why most AI apps feel half-built, what makes an app truly agent-native, and what your day-to-day experience looks like as a result." --- # What Is Agent-Native? Agent-native is a way of building software where the AI agent and the product surface around it are **equal partners**. The usual path starts with chat, adds typed actions, renders structured results inline, and grows into durable app pages around the same SQL state. The important part is that agents and humans share the same actions, database, and state. If you only remember one thing from this page, remember this: most AI apps today stop one step short of being useful, and that gap is the biggest mistake in the space right now. ## What it looks like as a user {#what-it-looks-like} Picture a background worker, inbox, calendar, form builder, or analytics dashboard. Often the first screen is chat: you ask what you want, the agent guides setup, shows a table or chart, and opens the right app view. Sometimes chat is docked on the right side of a full application. Sometimes there is no custom screen yet because the workflow is a scheduled job, queue, script, or external-agent integration. Across those shapes, you can: - **Start with the real operation.** One durable action can run from the CLI, HTTP, MCP, A2A, the app-agent loop, and later a UI. - **Click anything you'd normally click when there is a UI.** All the buttons, lists, dashboards, keyboard shortcuts — they all call the same operations the agent can call. - **Or just ask.** Type "reply to the email from Sara saying I'll be there by 3" into the agent. It opens the right thread, drafts the reply, and shows it to you for approval — exactly as if you'd done it by hand. - **See what it sees.** Open an email, and the agent knows which one. Select a chart, and the agent knows which chart. Highlight a paragraph and hit Cmd+I, and the agent acts on just that paragraph. - **Watch it work.** As the agent does things — opens views, edits drafts, runs reports — the UI updates in real time. You can stop it, redirect it, or take over with the mouse at any moment. - **Steer it like a teammate.** Give feedback, queue another task, edit its instructions, audit what it did yesterday. It remembers, and it gets better at your workflows over time. That's the experience agent-native is designed for. Now here's why most products don't get there. ## Why most "AI apps" fall short (The Ladder Principle) {#the-ladder} There's a progression most teams climb, much like a ladder, and most stop one rung too early. ### Rung 1 — a single LLM call (the anti-pattern) {#rung-one} A text box sends a prompt, the AI returns a string, and you display it. Maybe with a spinner. There's no way for the user to course-correct, no way for the AI to take action, no way to see what happened or why. You see this everywhere: "AI features" that are basically a "Summarize" button bolted onto a SaaS product. They look impressive in demos and break the moment reality gets messy. That's not a product; that's a toy. ### Rung 2 — a chat with tools {#rung-two} Now the AI can _do things_. It has tools — "draft email," "search contacts," "run query" — and a chat interface where it works in front of you, showing tool calls and results as it goes. This is what Claude, ChatGPT, and Cursor look like under the hood. This is a real step up. But on its own, it's still a chat window. There's no proper UI. No dashboards, no lists, no forms, no keyboard shortcuts, no team collaboration. If the AI gets confused, you're stuck retyping rather than just clicking the right button. Non-developers struggle to get real work done in this format. ### Rung 3 — agent + UI as equal partners {#rung-three} This is agent-native. You add a real, full-featured app around the agent — and crucially, every action the agent can take is also a button in the UI, and every button the user clicks runs the same logic the agent uses. One implementation, two ways in. Three things change when you reach rung 3: - **You stopped adding buttons to a chatbot. You added an agent to an app.** That's a much higher-quality product on both sides. - **The agent has real context.** It sees what you're looking at, what you've selected, what you just did. It writes to the same database the UI reads from, so its work shows up immediately. - **External agents can use it too.** Other agent-native apps can call this one's actions over the [A2A protocol](/docs/a2a-protocol). Claude Code, Codex, ChatGPT custom MCP apps, Cursor, and other MCP hosts can drive it as an [MCP server](/docs/mcp-protocol). One app, many entry points. That's rung 3. That's agent-native. ```html
Rung 3 · agent-nativeAgent + UI as equal partnersOne action surface. Every agent tool is also a button; every button runs the same logic the agent uses.
Rung 2A chat with toolsThe agent can act — but it is still just a chat window. No dashboards, lists, or shortcuts.
Rung 1A single LLM callPrompt in, string out. Impressive in a demo; breaks the moment reality gets messy.
``` ```css .diagram-ladder { display: flex; flex-direction: column; gap: 14px; } .diagram-ladder .rung { display: flex; flex-direction: column; gap: 6px; padding: 16px 18px; } .diagram-ladder .rung-2 { margin-inline-end: 48px; } .diagram-ladder .rung-1 { margin-inline-end: 96px; } ``` See [Key Concepts — Protocols](/docs/key-concepts#protocols) for how all of this hangs off the same action definition. ## Why every agent needs a UI {#why-every-agent-needs-a-ui} Even when the agent does all the heavy lifting, humans still need to: - **See what it's doing** — progress, intermediate output, what it touched - **Steer it** — give feedback, interrupt, queue the next task - **Manage it** — edit its instructions, skills, memory, scheduled jobs, connected accounts - **Inspect its work** — review drafts, audit history, roll back mistakes - **Share its output** — dashboards, reports, forms, links to send to teammates At minimum, "a UI for the agent" is an observability and management dashboard. At maximum, it's a full SaaS app with the agent embedded as a co-pilot. Both ends count as agent-native, and the surface can grow from one without a rewrite. You don't have to choose every surface up front. The agent can start in rich chat, render native inline UI, live inside a full application, or run from automation and external-agent channels around the same action surface — see [Agent Surfaces](/docs/agent-surfaces) for the concrete shapes and APIs. ## Why every app benefits from an agent {#why-every-app-benefits-from-an-agent} The flip side is just as important. Existing SaaS products keep hitting the same wall: 80% of what you need works great, and 20% you just can't change. Adding a chat sidebar rarely fixes that — the chat usually can't actually _do_ the things the UI can. Agent-native flips that. Because every action in the app is defined once and exposed as both a button and an agent tool, the agent can do everything the buttons can — and more — without a separate "AI world" to maintain. Natural language becomes a first-class input alongside clicks. The argument isn't "agents replace UI." It's "**agents belong inside applications, with a UI on top, as equal partners**." Even an automation-first app still needs some surface for humans to supervise, configure, and steer it — usually chat, a job history, or a small status page. See [Agent Surfaces — Automation-first app](/docs/agent-surfaces#headless). ## Agent + UI parity {#agent-ui-parity} This is the defining principle. - **From the UI** — click buttons, fill forms, navigate views. The UI writes to the database; the agent sees the results. - **From the agent** — natural language, other agents via A2A, Slack, Telegram. The agent writes to the database; the UI updates automatically. A good mental model is a shared state machine. The user surface and the agent surface both submit typed actions; those actions behave like reducer events with a schema and permission boundary around them: validate the input, authorize the actor, apply the state update, then broadcast/render the result. The UI is mostly a projection of that resulting state. The agent is not guessing from the DOM or mutating private component state directly; it interacts through the same typed action layer the UI uses. ```html
User surface
Open Design tab
navigate.panel
Add text block
plan.addBlock
Remove summary
plan.removeBlock
intent -> actionaction <- intent
Typed actionnavigateactor: user or agent · input: panel = Design
validateauthorizeapply update
Shared stateSQL + app state{ panel: "Design", lastActor: "user" }broadcast over live sync
state -> UIcontext -> agent
Agent surface
Add diagram block
plan.addBlock
Remove stale block
plan.removeBlock
Focus review view
navigate.review
``` ```css .diagram-machine { display: grid; grid-template-columns: minmax(150px, 1fr) minmax(300px, 1.45fr) minmax( 150px, 1fr ); gap: 12px; align-items: stretch; } .diagram-machine .surface { display: flex; flex-direction: column; gap: 8px; padding: 14px; } .diagram-machine .surface .diagram-pill { align-self: flex-start; } .diagram-machine .diagram-node { display: flex; flex-direction: column; gap: 2px; padding: 10px 12px; } .diagram-machine .core { display: flex; flex-direction: column; justify-content: center; gap: 8px; min-width: 0; } .diagram-machine .lane-row { display: flex; justify-content: space-between; gap: 8px; flex-wrap: wrap; } .diagram-machine .lane-row.bottom { justify-content: center; } .diagram-machine .diagram-arrow { font-size: 22px; line-height: 1; text-align: center; } .diagram-machine .diagram-panel { display: flex; flex-direction: column; gap: 8px; min-width: 0; padding: 14px; } .diagram-machine .steps { display: flex; gap: 6px; flex-wrap: wrap; } .diagram-machine code { align-self: flex-start; max-width: 100%; white-space: normal; } @media (max-width: 760px) { .diagram-machine { grid-template-columns: 1fr; } .diagram-machine .lane-row { justify-content: flex-start; } .diagram-machine .lane-row.bottom { justify-content: flex-start; } } ``` When the agent creates a draft email, it appears in the UI. When you click "Send," the agent knows it was sent. There's no separate "agent world" and "UI world" — it's one system. See [Key Concepts](/docs/key-concepts) for the architecture that makes this work. ## Customization usually reserved for power tools {#workspace-customization} The reason tools like Claude Code feel so powerful isn't the model — it's the **customization layer**: per-project instructions, skills, memory, sub-agents, connected services. You can shape the agent to your codebase, your preferences, your team. Agent-native gives every user that same customization layer — without ever leaving the app. Each app comes with personal **agent resources** that you (or anyone on your team) can use to: - Edit team-wide rules everyone's agent reads - Let the agent remember preferences automatically as you correct it - Write reusable how-to guides as `/slash` commands - Keep custom sub-agents for specific tasks (invoked with `@mentions`) - Schedule jobs to run on a cron (e.g. "every Monday morning, summarize last week") - Connect external services (Gmail, Stripe, Slack, internal APIs) via per-user MCP servers The twist: it's all stored in the database, not the filesystem. There's no dev environment to spin up, no container per user. Every user gets their own full set of agent resources — personal memory, personal connections, personal skills — for essentially free, because it's all rows in a table. That's what makes Claude-Code-level flexibility viable inside a real multi-tenant SaaS product. See [Agent Resources](/docs/agent-resources) for the full concept. ## What makes it different {#what-makes-it-different} | Approach | Description | | -------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------- | | **Traditional apps with AI bolted on** | The AI is an afterthought. Limited to autocomplete, summaries, or a chat sidebar that can't actually do anything in the app. | | **Pure chat / agent interfaces** | Powerful but inaccessible. No dashboards, no workflows, no persistence. Non-developers can't use them effectively. | | **Claude Code / Codex for SaaS** | Great for devs on their own machines. Doesn't translate to multi-tenant SaaS — one codebase per user on a dev-box doesn't scale. | | **Agent-native apps** | The agent is a first-class citizen. It shares the same database, the same state, and can do everything the UI can do — and vice versa. | ## Whole-team development {#whole-team-development} Agent-native isn't just for developers. Because the agent can edit the app's own code, evolving an app stops being a developer-only activity: - **Designers** update designs directly in the running app through the agent - **Product managers** add functionality and update flows by describing them - **QA** tests the app and asks the agent to fix what's broken - **Anyone on the team** contributes through natural language The vision: fewer handoffs, one person doing the work of a small team. ## Start from a template and customize {#fork-and-customize} Agent-native apps follow a start-from-a-working-app model. You start from a **template** — Calendar, Content, Slides, Analytics, Mail, Clips, Design, Forms, Dispatch — and make it yours. Each one is a complete, working SaaS product, not a blank scaffold: 1. Pick a template on [agent-native.com/templates](/templates) 2. Use it immediately as a hosted app (e.g. mail.agent-native.com) 3. Create your own app from the template when you want to customize — "connect our Stripe account," "add a cohort chart" 4. The agent modifies the code to match your needs 5. Deploy your app to your own domain — or stay on agent-native.com Because it's _your_ app, not shared infrastructure, the agent can safely evolve the code. Your app keeps improving as you use it. See [Templates](/docs/cloneable-saas) for the full story. Not ready to start from a whole template? You can also try agent-native by adding a **skill** to a coding agent you already use — install the Plans skill with `npx @agent-native/core@latest skills add visual-plan`. See the [Skills Guide](/docs/skills-guide#app-backed-skills). ## Composable agents {#composable-agents} Agent-native apps can talk to each other. From inside the mail app, you can tag the analytics agent to query data and include the result in a draft email. The agents discover what other agents are available, hand off work between each other, and surface the results in the UI you're already in. This is powered by [A2A](/docs/a2a-protocol) and [MCP](/docs/mcp-protocol) under the hood — same definition, multiple surfaces — but as a user, all you have to know is "I can ask any of my apps for help with anything any of them can do." ## What does this look like in code? {#what-does-it-look-like-in-code} If you're building or extending an agent-native app, here's the central pattern: every operation in the app is an **action** — defined once, available to both the agent and the UI. {\n const db = getDb();\n await db.insert(schema.replies).values({ emailId, body });\n },\n});' } annotations={[ { lines: "7", label: "Tool surface", note: "The `description` is what the agent reads to decide when to call this as a tool.", }, { lines: "8", label: "Typed contract", note: "One zod `schema` validates input from **every** surface — agent, UI, HTTP, MCP, and A2A.", }, { lines: "9-12", label: "One implementation", note: "The `run` body is the single source of truth. The UI button and the agent tool both execute exactly this. `getDb()` and `schema` come from your app's `server/db/index.ts` — see [Database](/docs/database).", }, ]} /> ```tsx // In any React component — same action, called from a button const { mutate } = useActionMutation("reply-to-email"); ; ``` ```tsx // And the agent panel mounted anywhere in your app import { AgentSidebar } from "@agent-native/core/client/agent-chat"; ; ``` One action, many surfaces: the agent calls it as a tool, the UI calls it as a typesafe mutation, [native chat](/docs/native-chat-ui) can render explicit widget results, external agents reach it over [A2A](/docs/a2a-protocol), and MCP hosts call it through the app's [MCP server](/docs/mcp-protocol), optionally with MCP Apps UI resources and standard remote MCP OAuth handled by the framework. See [Actions](/docs/actions) for the full reference. ## What's next {#whats-next} - [**Getting Started**](/docs/getting-started) — start with chat, add an action, render a native result, and grow into a page - [**Agent Surfaces**](/docs/agent-surfaces) — choose chat, native inline UI, full app pages, embedded sidecar, automation, or external agent access - [**Key Concepts**](/docs/key-concepts) — the architecture: SQL, actions, polling sync, context awareness, portability - [**Templates**](/docs/cloneable-saas) — templates as complete products you own - [**Agent Resources**](/docs/agent-resources) — the per-user customization layer (skills, memory, instructions, MCP) backed by SQL, not files - [**Dispatch**](/docs/dispatch) — the workspace control plane: secrets vault, Slack/email inbox, cross-app delegation - [**Extensions**](/docs/extensions) — sandboxed mini-apps the agent creates instantly without code changes - [**Drop-in Agent**](/docs/drop-in-agent) — mount `` into any React app - [**FAQ**](/docs/faq) — quick answers on cost, hosting, models, and templates