Why AI Agents Don't Scale In Companies - According To The Engineer Who Built Claude Code

Most organisations stall at the same place: they add more AI and nothing compounds. The engineer who created Claude Code published a five-level map of why, and the answer at every level is organisational, not technical.

Not one of the five bottlenecks is the model. Which is why buying more capacity moves you faster inside the level you're already stuck on, and no further.

Why do most companies fail to scale AI agents?

Because the constraint was never throughput. Every level of adoption has exactly one thing holding it there, and all five are properties of how the organisation runs - approvals, attention, review capacity, trust, guardrails.

This matches what the enterprise research keeps finding from the other direction: AI agents don't overcome weak operating models, they reveal them. Layer agents onto legacy workflows and the bottleneck relocates rather than disappearing.

What the map below adds is a named, first-party version of that, from someone with an unusual vantage point - and a specific claim about where his own employer sits.

What are the five levels of AI adoption?

Boris Cherny - who created Claude Code - published Steps of AI Adoption on 16 July 2026. The structure, and the part almost every summary leaves out, is the fourth column:

Read down that last column. Approval processes. Attention. Review capacity. Trust. Guardrails. Every one is something about how an organisation is run. Not one is a property of the model.

That is the whole argument, and it's sitting in the source that everyone quotes.

Is Anthropic at level 4?

No. Anthropic operates at step 3 across the organisation. Cherny personally claimed step 4 on 17 July 2026, the day after publishing.

This is the most useful fact in the entire document and it is almost never the headline. The company that builds the tool has not reached the top of its own creator's map. Step 4 means a thousand or more agents running at once — the constraint isn't licences or budget, it's that operating at that scale requires guardrails an organisation has to actually build.

If Anthropic is on three, "get to level four" is not a realistic next-quarter objective for anyone. The useful question is which single level you are stuck below, and what specifically is holding you there.

Where do the "2% at level 3" and "0.1% at level 4" figures come from?

🚨 Not from the source. Those percentages are circulating in video coverage of the framework and do not appear in Cherny's table. They originate with the creators covering it.

I checked because I nearly used one. If you have seen those numbers presented as research, they are not — there is no published population survey behind them.

The real number is more useful anyway: level four means 1,000+ agents. That is a concrete threshold you can measure yourself, rather than a percentile you cannot verify.

Does spending more on AI help you scale?

Because money is not what any of the five bottlenecks are made of.

Look at what actually changes between adjacent levels:

• 0 → 1 is an approval, not a purchase. Work built on a personal account and a personal card can never be handed to anyone else, which is what keeps that level stuck.
• 1 → 2 is the moment you stop being the only thing checking the work.
• 2 → 3 is when output stops queueing behind one person's reading speed.
• 3 → 4 is when work starts without being asked for, and the organisation can live with what happened overnight.

Each of those is a decision, a permission, or a habit. You can buy more capacity at any level and stay exactly where you are — which is precisely why spending more feels like it ought to work, and doesn't. You are paying to move faster inside a level you never left.

There is a specific thing to turn on at each one. That's the part I put on a separate page, linked at the bottom.