---
schema: 1
type: term
title: '**agent-native architecture**'
description: >-
  An approach to software design in which an app's own capabilities are exposed
  to AI agents, not just to human users — captured in three levels: anything the
  user can do the agent can do, anything the code can do the agent can do, and
  anything the developer can do the agent can do (e.g., an agent submitting a
  pull request in response to a feature request).
canonical: >-
  https://liquidgold.cc/channel/everyinc/explore/term/agent-native-architecture-9d537a8502c0
markdown: >-
  https://liquidgold.cc/channel/everyinc/explore/term/agent-native-architecture-9d537a8502c0.md
llms: null
parent: 'https://liquidgold.cc/channel/everyinc/llms.txt'
channel: everyinc
updated: '2026-10-23T22:12:35.346Z'
language: null
children: []
---


# \*\*agent-native architecture\*\*

Canonical: https://liquidgold.cc/channel/everyinc/explore/term/agent-native-architecture-9d537a8502c0

Parent: [Channel index](https://liquidgold.cc/channel/everyinc/llms.txt)

## Definition

An approach to software design in which an app's own capabilities are exposed to AI agents, not just to human users — captured in three levels: anything the user can do the agent can do, anything the code can do the agent can do, and anything the developer can do the agent can do (e.g., an agent submitting a pull request in response to a feature request).

In the channel's material, **agent-native architecture** is a concept Dan Shipper introduces in his 2026 software predictions: the idea that **agentic development should live inside the app you're actually using** — not just in Discord, a terminal, or a separate tool. It's defined by three escalating levels of agent access to an app's own capabilities:1. **Anything the user can do, the agent can do.** Most apps today fall short here — in Cora, for example, many things you can do in the app the chat agent can't. Once an agent can change your settings or manage your account on request, an app becomes partially agent-native.2. **Anything the code can do, the agent can do.** Actions that exist in code but that users can't trigger themselves — like regenerating an entire email brief — should be agent-accessible, triggered by user or developer.3. **Anything the developer can do, the agent can do.** You ask for a customization; the agent submits a pull request — a bug fix for everyone, or your own custom version of the app.- Dan credits **Anthropic** with nailing this — Claude Code, where “anything you can do on your computer, Claude can do” — and notes **Notion** is close, treating agents and humans as **first-class citizens** rather than pasting agents into human-collaboration software.- **Every is experimenting at level three**: users can DM a tool or mention it in Discord to request a feature or flag a bug, and it writes code and submits a PR. The next step is doing this inside the app itself.- In Demo Day coverage of Every's Think Week, a non-coder rebuilt a taste-analysis app on top of compound engineering and agent-native architecture — an example of the architecture opening building to people who don't write code.- Related framing from Mike Krieger's conversation: understanding **what an agent-native app looks like** is treated as part of rethinking what's possible with agentic models.The prediction is that by the end of 2026 this pattern will be **broadly distributed** — agentic development living inside the products people use daily, reshaping how features get built and requested.- **Agent-native architecture** = agents operating inside the app itself, across three levels: user actions, code actions, and developer actions.- **Anthropic (and Notion)** are cited as current leaders in treating agents as first-class citizens.- **Every** already practices level three via Discord-triggered pull requests.- It's predicted to become widespread in software by the end of **2026**.