It Spreads Sideways. Someone Still Has to Light It.
Anthropic's Claude Code lead published a five-rung adoption ladder this week. Microsoft published the measurement fifteen days earlier, and the two do not agree about the size of the prize.
On July 16, Boris Cherny published a table. He helped build Claude Code at Anthropic, he talks to engineers at other companies most days, and he says he keeps hearing the same story: one person is 10x'ing their output with Claude, and the rest of the org hasn't caught up. So he mapped what he sees. Five rungs, zero through four.
Eight hours before that post, Amjad Masad, who runs Replit, described something from inside his own company. The same engineers had 3x'd their output in six months. Support was resolving its hardest tickets 60 percent faster. He has a name for the shape he thinks he is watching: the self-driving company.
Fifteen days before either of them, three researchers at Microsoft published a number.
They studied tens of thousands of engineers through the company's early-2026 rollout of Claude Code and GitHub Copilot CLI. Four months, real telemetry, no survey. Engineers who adopted the tools merged about 24 percent more pull requests than they otherwise would have, and the lift held steady across the entire window.
Ten times. Three times. Twenty-four percent.
The three are not counting the same thing, and the gap between them is the story. Cherny is reporting what he hears. Masad is describing a company that sells AI coding tools and is staffed by the people who build them. Microsoft counted merged pull requests and told you exactly what it counted, including that "a merged PR is not the same as the value it delivers."
The number falls as the instrument sharpens. Loose quantities are enormous. Precise ones are modest.
What the ladder gets right
The table is serious work, and the criticism only lands if you take it seriously first.
Each rung gets a role, an agent count, a bottleneck, and guardrails. Step 0 is Gated: no access, legacy approvals. Step 1 is Assisted, one agent, you and it as a pair. Step 2 is Parallel, about ten agents, and you become an Orchestrator. Step 3 is Supervised autonomy, roughly a hundred agents, and Cherny calls the role Manager of managers. Step 4 is AI-native, a thousand or more, steering by intent.
Now read the bottleneck column straight down, ignoring everything else.
Your attention. Reviewing output. Trust in the loop and your team's decision throughput. Identifying and automating work at scale.
The model is not in it. Not on one rung. The person who helped build the product, published by the company that sells the model, put out an adoption ladder in which the model is never the thing standing in your way. Every constraint on that table is a human quantity. I have been making this argument for months and I have never had a cleaner statement of it than the one Anthropic just published by accident.
Every constraint on that table is a human quantity.
The mess is not a side effect
Arseny Kapoulkine, who spent years as a technical fellow at Roblox and wrote the tools a lot of the game industry runs on, answered Cherny in one sentence:
"one person is 10x'ing their output with Claude while the rest of the org is busy dealing with the resulting mess"
Several hundred likes, which in this corner of the internet is a room nodding.
Cherny says the org hasn't caught up. Kapoulkine says the org is not standing still. It is cleaning.
That is the part the table leaves out. Output has to land somewhere, and an organization can only absorb so much of it. The limit is not the model's. It is how fast people can read what came out, decide about it, and put their name on it. Push past that line and the extra has two places to go, and neither of them is up.
It queues, and the person having the best month of their career watches it go stale in a backlog. Or it ships unread and becomes what one of Cherny's own readers calls slop. The first teaches your most motivated engineer that speed is pointless. The second teaches everyone downstream that the new work cannot be trusted. A wall or a mess. Pick one.
His readers know exactly where their own line is. One of them no longer reads his code at all and still runs "only 1-4 sessions in parallel because that is my speed of verifying." These are not skeptics. They are enthusiasts at the top of the curve, describing a wall made of their own eyes. A self-driving company whose drivers say they cannot stop watching the road.
Cherny's table concedes it in the step 3 row. The trap, it says, is "scaling agent count before the loop has earned widespread trust." The agent count is the output, not the input. You do not climb by adding agents. The agents show up when something else has already changed, and the thing that changes is how much a person is willing to stop looking.
How it actually moves
Back to the Microsoft paper, because it answers a question the table does not ask: how does any of this spread in the first place?
Not by memo. First use, the authors write, "spread primarily through social networks," and their recommendation to any org attempting this is to treat "visible peer use as central to rollout strategy."
They put numbers on it. An engineer had 54 percent higher odds of trying the tool where a quarter or more of the people they trade code reviews with had already used it. An engineer whose skip-level peers, meaning the engineers who share their manager's manager, were largely using it had 216 percent higher odds. That was the strongest signal in the study.
Sideways. Through the people you already trade work with, and not down through a mandate or a literacy program that 85 percent of employees say they cannot apply to the job they actually do.
Anyone who has watched a transformation program die should find that encouraging. The spread does not have to be installed. It installs itself, through the same social wiring the org already runs on.
Except it needs a light.
The manager number
Microsoft modelled the exact variable. In the paper's own words, a binary indicator of whether engineer i's direct manager used Copilot CLI.
"An engineer whose manager used Copilot CLI had higher odds of both trying it (+82%) and, more modestly, sticking with it (+22%)."
A manager putting their own hands on the tool nearly doubles the odds their reports try it.
Then the other half. On whether managers picked it up themselves, the paper reports that they "looked no different from the reference." No more likely than a mid-level individual contributor. No less.
The one person whose adoption moves everybody else's is no more likely than anybody else to adopt.
So the two findings stop competing and become one machine. The manager's crossing is the ignition. The peer network is the amplifier. Eighty-two percent lights it, two hundred and sixteen carries it, and the reason most orgs have neither is that nothing struck the match.
What spreads on its own
Grassroots adoption works, and it does not need a leader. Look at what it produces when it doesn't have one.
Most enterprises have now found an agent running that nobody signed off on. Gartner projections reported this month have the average Fortune 500 carrying more than 150,000 agents by 2028, up from fewer than fifteen in 2025. The same projections have 40 percent of enterprises demoting or decommissioning agents by 2027, after a production incident.
That is not a step 3 organization. That is a step 0 organization full of step 2 individuals, each one locally optimized, none of it adding up.
Which is Cherny's opening sentence, at scale. Unseeded sideways spread does not disprove his ladder. It manufactures the exact problem he opens with.
Everyone is naming the same top rung
Masad calls it the self-driving company. Cherny's step 4 is AI-native, a thousand agents, the human steering by intent. Back in May, Brian Armstrong told Coinbase he was "rebuilding Coinbase as an intelligence, with humans around the edge aligning it," and cut about 700 roles on the way.
Three serious people, ten weeks, the same shape: the organization drives, and the human moves to the edge.
I think the edge is the wrong place to stand, and the evidence in this piece is why. The enthusiasts running the most agents say they cap at four because of their own eyes. Cherny's own step 3 says the trap is scaling agent count before the loop has earned trust. His entire bottleneck column is attention, review, and trust. Every one of those is a statement about a human being close to the work, not at the edge of it.
Armstrong's memo actually contains the refutation. A few lines under the sentence about humans at the edge, he mandates the opposite: "No pure managers. Every leader at Coinbase must also be a strong and active individual contributor." Player-coaches, he says. Hands dirty. That is the thesis of the book I just published, arrived at independently and written into HR policy instead of a chapter.
You cannot be at the edge and have your hands dirty. He is right the second time.
Teach the teachers
I ran a version of this at a Fortune 500 pharma, across a coding-agent rollout to a large technical organization, and the pattern that worked was not a program.
It was: cross first, personally. Build something real in your own domain, badly, and then less badly. Then teach the handful of people closest to you, not by presentation but by showing them the thing and the scars on it. Those people teach the people next to them. The distance the idea travels from you is short. The distance it travels after you is the whole org.
You are not the distribution channel. You are the ignition source, and then you get out of the way of a network that moves faster than you can. One is you. The other is what happens next.
The two objections
Two arguments cut against this, and both deserve better than a wave.
You can buy it. The day after that Microsoft paper went up, Microsoft stood up a $2.5 billion, six-thousand-person company whose entire job is making enterprise AI deployments work. Its pilot cut a supply-chain Copilot rollout from fourteen months to five. Aaron Levie thinks the implementation work ahead "will exceed anything we imagine today." If capability can be imported wholesale, then the ceiling is your budget and your vendor list, not whether anyone senior has personally built anything.
Governance may matter more than any leader's history. A survey of 157 enterprises found half had shipped an agent that passed internal evals and then failed in front of a customer. Only 5 percent fully trust automated evaluation. Sixty-six percent are engineering toward zero human in the loop anyway. Their phrase for it is that the autonomy is arriving faster than the assurance, and none of it cares who approved the deployment.
Both are true, and I am not waving at the second one. I spent a whole piece ten days ago arguing that in a regulated industry the approval queue, not the keyboard, is what actually holds you back: AI has eaten the keyboard, it has not eaten the queue. Governance is necessary. It is not sufficient, for a small and unglamorous reason: someone has to be able to tell whether the governance is pointed at anything real. A person who has never run the loop cannot tell a working verification stack from a slide that says Verification. They will approve the slide. They approve the slide constantly. That judgment is not a policy you can buy.
The rung you set
Cherny's ladder is not wrong. It is aimed.
It is written for the engineer climbing it, which is the right audience for the person who built Claude Code, and it is useful if that is you. The person it never addresses is the one who decides how far it goes. Look at his own step 0, the bottom rung, the one whose stated bottleneck includes a "lack of true technical voices in decisionmaking." The exit condition he lists is executive and buyer alignment.
His table already says the way off the floor is a person with authority. It just doesn't say that person has to have used the thing.
None of this means an organization can't evolve without you. It plainly can, and most enterprises are proving it in the shadows right now. It means something narrower and more uncomfortable: if nobody senior ever crosses, the org can stall at the rung where its leaders stopped, and what it accumulates instead of capability is mess. No published adoption ladder I found this week models falling. Every one of them models climbing.
You already know which rung you're on. The number you don't have is the other one: what your reports' odds would be if you were on it.
Sources
Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI (Murphy-Hill, Butler, Savelieva, July 1, 2026). Tens of thousands of engineers, four months, the 24 percent lift and the social-exposure numbers.
Boris Cherny, Steps of AI Adoption (July 16, 2026) and the post announcing it.
Amjad Masad on Replit's six months (July 16, 2026), where the 3x, the 60 percent, and the self-driving company come from.
Arseny Kapoulkine's reply (July 16, 2026).
Microsoft's $2.5 billion, 6,000-person AI implementation unit (CNBC, July 2, 2026).
Why AI readiness training fails (HR Dive, on Docebo's 2026 AI Readiness Gap report, 2,000 respondents across six countries).
Related, from me: The Harness Is the Moat on why the model is the commodity layer, Ep. 1: Two Groups on the population split underneath all of this, and You Don't Have to Write the Code on what Anthropic's 400,000-session study found actually predicts success.
Justin Johnson writes Run Data Run. His book on crossing this particular gap is Builder Leader (builder-leader.com).