Boris Cherny, creator and head of Claude Code, keeps seeing the same four steps as teams adopt AI.
His opener: one person is 10x'ing their output while the rest of the org hasn't caught up.
He mapped the climb — the tweet, the artifact.
Each step names your role, an agent headcount, the thing that unlocks, and the bottleneck that gates the next rung.
The headcount is exponential: 0, one, ten, a hundred, a thousand-plus.
Your job title mutates along the way.
All table content from the artifact, condensed. The "products that help" column is omitted here on purpose — it's an all-Anthropic shopping list, and that's an on-air talking point, not a diagram.
The thread's core claim: there's no one right path, and spending more on tokens doesn't advance you a single rung.
Progress is a loop — find the bottleneck, build the guardrail that breaks it, climb, repeat.
Every bottleneck on the table is a human quantity: attention, review, trust, decision throughput.
The model never appears in that column.
"The trap is scaling agent count before the loop has earned widespread trust." — step 3's bottleneck, and arguably the whole episode in one sentence.
Usage dashboards measure "activity, not return."
The better question: "would you have spent engineering effort on this anyway?"
"If yes, how much and what would it have cost in manual eng-hours? That's your return."
Thariq Shihipar says Anthropic removed over 80% of Claude Code's system prompt for the newest Claude models, with no measurable coding-eval loss.
That rhymes with Boris's ladder: better models need fewer permanent rules and better context boundaries.
The new advice is progressive disclosure — smaller skills, lighter CLAUDE.md files, and context loaded when the model actually needs it.
Boris closed the thread with a question built for a podcast: "Anthropic is on step 3 and pushing toward 4."
"Personally, I just hit level 4. Curious where you are -- what step is your team on?"
Answer five questions live.
Argue with the verdict.
Then ask the audience for theirs.
1.3M views brings every kind of reader.
The replies split into skeptics with real objections, rival taxonomies, and — deliciously — bug reports about the artifact itself.
Tap the amber buttons for Boris's answers.
"I've seen people who I expected to adopt AI almost actively avoid it... There is clearly a camp that 'gets it' and another apathetic group with no interest."
the ladder assumes you want to climb"Great, but i think agents do make mistakes, in that case, how does this look like? backwards and keeps moving or does it require additional support?"
where's the down escalator?"Would genuinely love to take a course on moving from level 0 to 'AI Native'... Literally don't even have a mental model of what this would look like irl."
the demand side"I call the point where you stop reading code and just live in the agents 'AI Native,' and the point where PRs become continuous 'AI Industrialized.'"
"I just have one question: Who added 'LOAD BEARING' to Fable? 😅"
"This tread looks like it was made by Claude"
"Page not scrollable on safari on Mac" · "This doesn't load on android" · "scroll issues / buggy artifact"
the replies became a bug tracker"this is great, but Claude can't crawl your artifact... claude should be able to crawl a public artifact?"
— with a screenshot of Claude failing to fetch the page.
our prep hit the same wallThe manifesto for the thousand-agent future shipped as an artifact that wouldn't scroll on Safari, didn't load on Android, and couldn't be read by Claude itself.
Even prepping this segment, the automated fetch got an empty JS shell — a human had to open a browser.
The ladder is real; the rungs are still wet paint.
2026 is the year orgs moved past the lone 10x engineer and hit the org-wide adoption wall.
Five labs committed $8B+ to deployment teams this quarter, and surveys put the barrier at organizational readiness, not the model.
Boris's ladder is a map for exactly that gap — which is why it got bookmarked more than it got liked.