The AI-Native Company

What AI-native means

A software factory that learns and heals itself.

Agents do the core work. People work alongside them, steering and making the calls. Every lesson lifts the whole company.

How a software factory is built

Select a layer to see how it works.

The basic unit is one agent working in a loop, and a software factory is many of those loops running at once.1

Illustrative simulation, not real data

  1. 01The loop

    After Osmani's model

    One agent, one job, run until the result checks out.1

    Context · Act · Verify · Repeat

  2. 02The harness

    After Osmani's model

    Walls that keep a loop safe and decide when it is done.1

    Sandbox · Tools · Memory · Done checks

  3. 03The factory

    After Osmani's model

    Many harnessed loops at once, with people in charge.1

    Work queue · Harnessed loops · Review gate · Production · People in charge

    “It is not a bigger agent; it is an org chart made of loops.”

    Addy Osmani, On the software factory1

  4. 04The learning loop

    The author's addition, not Osmani's

    Every run leaves a lesson the next loop starts from.2,3,4

    Lesson · Company brain · Next loop · Self-healing

  • After Osmani's model
  • The author's addition
  • a person decides
  • automatic done checks
  • a failure and its fix
  • a failure in production the lesson is saved an agent drafts the fix a person releases it

Four layers, each built on the one before. The loop: one agent gathers context, acts, verifies and repeats. The harness: the loop inside walls of sandbox, tools, memory and done checks. The factory: many harnessed loops take work from a queue. They pass one review gate, where people decide, before production, with people in charge of the whole. The learning loop, the author's addition: every run saves a lesson in the company brain that the next loop starts from.

  1. The loop. The agent collects what it needs, does the work, tests the outcome and repeats until the job is done. Everything bigger is built from these loops. (After Osmani's model)
  2. The harness. The loop runs in a sandbox, with only the tools it needs, memory that lasts between runs, and checks that say when it is done. Without these walls, an agent is unsafe and may never stop. (After Osmani's model)
  3. The factory. Work arrives in a queue, many harnessed loops take it on at once, and every change passes one review gate before it goes live. People oversee the whole system and make the call at that gate. (After Osmani's model)
  4. The learning loop. Lessons go into the company brain, so the factory gets better every time it runs. This is how it learns and heals itself: agents catch failures and draft the fix, and a person releases it. (The author's addition, not Osmani's)

Illustrative simulation, not real data

Loop, harness and factory follow Addy Osmani's model in Software Factories, Light and Dark (20 Jul 2026), redrawn for this page. The learning loop is the author's addition, not his.1,2,3,4

The definition

An AI-native company runs like a software factory that learns and heals itself. Agents build in self-improving loops. People and agents work together in the same sessions. Named people make the decisions, and every lesson goes into a shared company brain. When it works, each new feature takes less time and money to build than the one before.

The shift

AI-assistedAI-native
Work moves down a line of handoffsWork runs in loops that save what they learn
Engineers write code with AI helpEngineers build and run the factory that writes it1
Each person works alone with an AI chatPeople and agents work in shared sessions
Each project starts from scratchEach project starts from past work
Memos argue for ideasWorking demos test them

What it is made of

Loops that heal

Agents watch issues, bug reports, alerts and customer questions. They catch failures, save the lesson and draft the fix. A person releases it.4

Multiplayer by design

People and agents work in the same sessions, from Slack, the web or any tool. Anyone can join, steer, or pick up where someone left off.5

A software factory

One agent per task, each with its own instructions and tests. Every change passes the company's standards as automated checks.4,6

One company brain

Every lesson lands in owned, tracked files that people and agents read and correct. Friday's fix is Monday's starting point.2,3,7

People at the gates

As coding speeds up, the bottleneck moves to human decisions. Named people set the standards, make the calls and release the work.8

Everyone AI-fluent

Every hire, in every function, works through agents. Fluency is judged by the work, not by usage counts.

Already working this way

StripeMinions

Engineers start these background coding agents from Slack. Over 1,300 code changes a week are written entirely by agents and merged after a person reviews each one (Feb 2026).9,10

RampInspect

A shared coding agent that runs tests, reads production telemetry and sends screenshots to prove its work. About 75% of merged code changes came from it by May 2026.5,11

RampGlass

An AI workspace where every session starts from what the company already knows. 323 employees made 6,500+ edits to that shared knowledge (Oct 2026).7

AnthropicClaude

Claude writes more than 80% of the code merged to production, a figure Anthropic calls conservative (May 2026). Merged changes rose from about 2,500 to 30,000+ a day by Sept 2026.12,13

OpenAICodex

A team that grew from 3 to 7 engineers built a product in five months with zero hand-written code, about a million lines (Feb 2026).14

ShopifyRiver

Its Slack agent works only in public channels, so every conversation becomes searchable. It co-authored 3,536 merged code changes in 30 days (May 2026).2

VercelAI SDK factory

One agent per task picks up issues on its open-source toolkit. Four weeks in, agents wrote 25 to 35% of weekly merges, and a person approves each (Aug 2026).4

EveryCompound engineering

Five in-house products, each built and run mainly by one person. Every bug and insight is written down for future agents to use (Dec 2025).3

NBERStudyindependent

An observational study of 500,000+ developers linked coding agents to about 30% more releases and 240% more commits. Count what ships (May 2026).15

What one person learns, the whole company knows.