Four steps from chatting to hands-off AI

“you move up by removing a bottleneck, not by adding more AI”

Boris Cherny built Claude Code at Anthropic, and in July 2026 he published a short map of how people adopt AI at work: four steps, from one supervised session to a system that starts its own work. Austin Marchese’s walkthrough translates the map for people who do not write code. The useful part is the diagnosis. Each step has one named bottleneck, and the bottleneck you feel tells you which step you are on and what to change next.

01Four steps, four bottlenecks

Cherny wrote the map after hearing the same thing from company after company: one person is getting ten times the output from AI while the rest of the organization has not caught up. The steps describe what that person does differently at each stage. Below the first step sits the place most people start, using AI only in a chat window, asking questions and copying answers out.

Step	What it looks like	The bottleneck
1 · Assisted	One AI session at a time does tasks for you; you review every change	Your attention: you read everything
2 · Parallel	Five to ten agents at once, each on separate work, checked by machine first	Steering: you still prompt and redirect each session
3 · Supervised autonomy	Agents start scheduled and background work on their own; you review results	Trust, your decision speed, and cost
4 · AI native	The system kicks off most work itself; you set direction and look at exceptions	Choosing what to automate, and keeping it cheap

Each bottleneck is the admission ticket to the next step: what blocks you at step one is exactly what step two removes. The rest of this guide takes the steps one at a time.

02Step 1: stop copy-pasting

Cherny describes step one as one person with one agent, mostly supervised: “a fast pair programmer” (1:33), an engineer’s phrase for a colleague working beside you on the same task. For everyone else it means an assistant that does tasks rather than answering questions.

Most people believe they are here already. Marchese offers two questions to check (2:01):

• Do you copy AI output and paste it into other tools?
• Do you copy information from outside the chat and paste it in?

A yes to either means you are still the courier. The fix is access: a tool that reads and edits files directly (Claude Code is the one Marchese uses, for far more than code), plus connectors that link the accounts where your information already lives, so the AI pulls what it needs and writes results where they belong. Delegate to your agent in stages covers connecting that first account and how much power to grant on day one.

Then you hit the wall Cherny names for this step: “you feel you must read everything, so you never look away” (3:28).

03Step 2: machines check first

Step two is running several agents at once, five to ten in Cherny’s version, each on its own piece of work. What makes that possible is agents that finish a task and check their own output, so you look at a final product instead of every intermediate step.

“a backlog that used to take the team weeks becomes one engineer's afternoon of orchestration”

Boris Cherny, “Boris Cherny's 4 Step Playbook to 10x Your AI Productivity” · 4:11

That speed only arrives once a machine does the checking, so the checking is the first thing to hand over. Marchese splits the checks into two kinds (4:38):

• Rule-based checks. Yes-or-no questions with no room for debate. For code, that means automated tests; for a report, the right fonts, the right colors, no banned punctuation. Write the list once and a machine can run it every time; Loop the agent until the work passes builds that loop.
• Taste-based checks. Quality calls with no objective answer: does the design look good, is the writing tight. Write your standards down and have a second AI judge drafts against them. Why AI makes slop and how to catch it covers what those written standards need.

The second move is permissions. An agent working unwatched needs room to act without the ability to cause permanent damage.

“thinking of agents like a drunk intern”

Austin Marchese, “Boris Cherny's 4 Step Playbook to 10x Your AI Productivity” · 6:10

If the agent can reach something breakable, assume it will break it, and remove the access rather than hoping. Claude Code’s auto mode is the preset version: low-risk actions run without asking, risky ones still stop for your approval (6:29). Assume your agent uses every tool it has explains why the boundary belongs in the permissions, not the prompt.

The third move is picking work that can run side by side. Marchese keeps unrelated projects in separate folders so each agent sees only its own context (8:38), and within one project he picks units that do not touch: different customer proposals, different scripts, different parts of an app. When one big task needs many hands instead, Split big AI tasks into steps and checks covers cutting it into independent pieces.

The bottleneck that remains: “prompting and steering the model as you juggle sessions” (9:34). Every piece of work still starts with you typing.

04Step 3: work you never assigned

At step three the agents start work themselves. Cherny’s description: maintenance and cleanup that used to wait for someone to find time now runs continuously in the background (9:45).

The objection Marchese hears most: “it is faster if I just do it myself.” Usually true, for that one task, and beside the point. When he hired his first two contractors, every video took longer for two weeks while he trained them; then it got faster, and then he was not involved at all (10:37). Setting up the system is slow today and drops toward zero. Doing it yourself is fast today and fast forever.

Three moves turn a supervised setup into one that runs without a kickoff from you:

1. Write the procedure once. Do the task with the AI, going back and forth until the output is good, then tell it: based on this conversation, create a skill that makes the process repeatable (11:52). The skill becomes a standard operating procedure any future run follows. Capture the model’s discipline is this move in full.
2. Put it on a schedule. A routine runs the skill unattended, every Monday at 8 a.m. or on whatever schedule you set (12:37). Automate a workflow as a Claude routine covers what an unattended run needs.
3. Give it a place to report. A channel you actually check (Slack, WhatsApp, Telegram) where every finished run posts its result (13:04), the pattern Buzz puts AI agents in your team chat covers.

An agent that proposes work beyond the scheduled kind needs more: written goals and fresh context on what is actually happening. Make your agents proactive covers what to write down.

Cherny’s bottleneck here has three parts: how far you trust the runs you no longer watch, how fast you can decide on what the agents send up, and whether tokens (the units AI use is billed in) are spent efficiently as usage grows (14:29). Step four answers all three.

05Step 4: pick what to hand off

At the last step, most work is started by the system itself. Cherny: hundreds to thousands of agents run, and “quarter-long migrations become a workflow you kick off and check on” (15:43).

“you steer by intent and monitor by exception”

Boris Cherny, “Boris Cherny's 4 Step Playbook to 10x Your AI Productivity” · 15:34

Getting there is two judgment calls rather than more setup.

Choosing the candidates. Marchese’s rule comes from hiring: if someone can do a task 80 percent as well as you, hire them. Same test here. Where 80 percent quality is good enough, automate end to end. Where quality is critical, keep a person at the checkpoints instead. What still pays when AI does the work is about the judgment that stays yours.

Capping the cost. One of Marchese’s clients ran proactive agents without guardrails and woke up to a single task that had burned over $3,000 in tokens, stuck in a loop, spending until he manually stopped it (16:23). Two caps prevent a repeat:

• Pin each job to the smallest model that handles it, instead of paying the biggest model’s prices for routine work. Stop hitting your token limit shows how to set the model inside a skill.
• Set a maximum number of iterations on any loop or routine, so a stuck run stops on its own. Guardrails for hours-long agent runs covers the fuller set of stop conditions.

The steps compound in one direction: each one hands the machine another piece of your attention, and the bottleneck you feel right now is the only one you need to work on.

Further reading

• Austin Marchese · Boris Cherny’s 4 Step Playbook to 10x Your AI Productivity, the walkthrough this guide draws on
• Boris Cherny · Steps of AI Adoption, the original map, announced on his X account