Overview
Anthropic published “The Making of Claude Code” — an oral history of the product, told through interviews with nineteen people who built it, covering 2021 through 2026.
Three passages stopped me cold.
The way the builders describe how they work today, and the structure they predict as “next” — a one-person atelier on the outside was already living in both.
What the Oral History Is
The interviews were recorded between February and May 2026. Co-founder Ben Mann, Claude Code head Boris Cherny, Head of RL Shauna Kravec and others trace the path from an internal tool called “clide” to Claude Code’s breakout.
The numbers alone are striking. Boris testifies that by winter 2025, 100 percent of his code was written by Claude Code — not a single line by hand. There’s a mention of 88 commits in a single day.
But what hit hardest wasn’t the numbers. It was the structure.
Resonance 1: “A Swarm of Twelve Claudes”
Shauna Kravec, who leads reinforcement learning at Anthropic, says this:
“I have my whole swarm of twelve different Claudes running around—reading documents, updating things, pulling from Slack.”
The person training the models works through a swarm of twelve.
Here, sixteen are running. AI agents with defined roles investigate, implement, review, and report back. This setup has been shipping real client work for months.
The workflow of someone inside the frontier lab and the workflow of a solo studio in Tokyo turned out to have the same shape.
Resonance 2: “Managing the Claudes’ Manager”
Igor Kofman of the Labs team predicts the future this way:
“We’ll get to the next level of abstraction, where you’re not managing a bunch of Claudes—now you’re managing the Claudes’ manager.”
That “next” is present tense here.
The sixteen agents aren’t a flat pool. A coordinating AI receives tasks, assigns them to the right agent, inspects the results, and reports back. The human’s job is reduced to talking with the manager and making final calls.
I won’t publish the implementation details. But the idea itself — running AI in a hierarchy — doesn’t require scale to reach. What it requires isn’t capital. It’s time lived with the tools.
Resonance 3: The React Lesson and the “No Category” Strategy
The testimony that lands from the farthest distance comes from Adam Wolff, Claude Code’s first manager, who previously worked on React:
“It was a logo and a brand and a feeling, much more than a computer science idea.”
By the time React crossed a million DAU, it had become something else entirely. A product at scale outgrows its founding idea and becomes a feeling. Wolff expects Claude Code to follow the same path.
What survives on the receiving end isn’t the feature set. It’s the emotion.
ANDOOR refuses to call itself a “production company” or a “creative boutique.” We dropped the category and design for a state where the work communicates through atmosphere. That decision comes from the same recognition: the maker’s taxonomy evaporates on arrival. Only the texture of the experience remains.
The lesson the creator of React extracted from product history matched an instinct we’d already built a brand on.
The Frontier Isn’t Measured in Headcount
One line from Ben Mann summarizes the whole read:
You have to build something that works 20 or 30 percent of the time now, so that when the next model comes out, it works 80 percent of the time.
Move into the unfinished tool early, and lie in wait for the next model. The Claude Code team did exactly that inside Anthropic. That’s why they won.
The same move is available from the outside. Touch the tool’s limits daily, get stuck, build workarounds, get rewarded when the next model lands. Company size has nothing to do with this loop.
Your distance from the frontier is set not by the size of your organization, but by the hours you’ve lived with the tools.
This connects to what I wrote about Nadella: you can’t outsource your learning. Only the accumulated history of judgments — earned by living inside the tools — compounds the moment a new model ships.
Closing
There’s a worn-out line: the best way to predict the future is to invent it.
This felt different. The future wasn’t something to predict or invent. It was something to move into early.
Reading the builders’ own testimony and finding you could check your answers against it — that alone made this primary source worth the read.
The full oral history is on Anthropic’s site. Anyone who works with AI should read it.