Boris Cherny, creator of Claude Code at Anthropic, described a ladder of AI adoption: from "locked down" to AI-native. We've climbed most of it the hard way ourselves. I'll lay it out as a route — with marks at the spots where almost everyone stumbles.
"We've brought in AI" means different things at different companies: one engineer's productivity has multiplied, while the rest of the team hasn't rebuilt its processes yet. Boris Cherny offers a scale that brings order to this mess. Below is a retelling in our own words, with examples from how we at MILAGPT are building a network of AI employees.
A familiar picture: the company bought access to a chatbot, wrapped it in approvals, routed it through a corporate gateway and cut it back to old models. Agents: zero. The result lives in a chat window and dies there.
The diagnosis is precise: the bottleneck here isn't technology, it's process. The company counts the cost of tokens instead of the cost of the result, and at the table where decisions are made there's no one who has actually worked with agents. The way out isn't technical but organisational: come to an agreement with those who pay, and tear down the blockers.
The first real step: a person gets an agent as a partner. An edit that used to take half a day gets done between meetings. But the work is synchronous — you sit and proofread every line, because the agent doesn't check itself.
What moves you up isn't a smarter model but a self-verification loop you trust.
Tests, builds, a linter, a run in a real environment — so that the agent checks itself before you see the result. Until this loop exists, adding more agents is pointless: you'll just multiply the amount of proofreading.
Here one engineer conducts 5–10 agents, each with its own isolated branch so they don't elbow each other. Automatic code and security review are on by default. You look at the final diffs, not the process. A backlog of weeks fits into half a day.
The turning point. The agent no longer waits for a task — it watches its sources itself (channels, mail, data) and opens work itself. The question "did you read the code?" gives way to "what context was the model missing, and how do we give it next time?".
The trap of this stage, one we stepped into too: scaling up agents before the self-verification loop has earned trust. You get confident reports about things the agent never double-checked. One rule cures it: the agent must prove the result with a run, not a story.
The loop is closed: most agents are launched by AI itself; the human sets the intent and steps in where something went wrong. A migration that used to take a quarter turns into a workflow you just launch and check now and then. Only a handful have got here — but the whole ladder is already visible.
The framework's best idea is about the metric. Don't count tokens. Count the manual equivalent:
The half-day edit, the weeks-long backlog, the quarter-long migration — those are the units of measurement. And the reverse filter: "would an engineer have done this at all?" If not, you're automating rubbish.
If you personally proofread every agent reply, you're at stage 1, and the next step isn't "more agents" but a self-verification loop. If the agents are working but you're the one carrying tasks to them, that's stage 2 — time to hand them the sources of context.
You can spend a year climbing this ladder, or do it faster with someone who has already climbed it. Our AI employees are a ready-made stage 3: an agent that finds work itself, checks itself and brings a result, not a process.
Take an AI employee for a test drive and look at your company from the next stage up.
MILAGPT · Tashkent
Want a digital employee like this for yourself?
Write to Mila: she replies in seconds, at any time of day or night.
Write to Mila →Tell her about your business in Telegram and she'll help from the very first conversation.
Get your own Mila →Share:Telegram