The creator of Claude code released a guide to the 4 levels of AI coding adoption.

X.com Boris Cherny @bcherny Steps of AI claude.ai 414 7/16/26 1.3M Share this post on X yet Relevant Boris Cherny @bcherny There's no one right team and Claude tokens aren't CREATOR OF CLAUDE CODE DROPPED A FOUR STEPS OF ADOPTION THAT YOU SHOULD ALL STUDY. One engineer, one agent, mostly supervised-a fast pair programmer. You run one session at a time and review almost every change before it merges. Unlock: A change that used to fill an afternoon becomes so finish betw from step 1 to with a re BASICALLY AI PAIR PROGRAMMING WHERE, YOU'RE PROMPTING SMALL CHANGES IN AND THEN MANUALLY REVIEWING THOSE CHANGES YOURSELF. One engineer orchestrates 5-10 agents at once, each on its own worktree or git checkout, jumping between them. Claude checks its own work-tests, build, lint, security scan-before you see it. Auto mode is always on. Automated code review and security review are on by default. Out multiplies diffs rathe keystroke backlog work star Claude wr code. Boris Cherny @bcherny I talk to engineers at other companies every day and hear the same thing: one person is 10Xing their output with Claude but the rest of the org hasn't caught up. Watching teams adopt AI, I keep seeing the same 4 steps. I mapped them out here: Claude Steps of AI claude.ai 414 7/16/26 1.3M Share this post on X yet Relevant Boris Cherny @bcherny There's no one right team and Claude tokens aren't Review hand instead of it, your Prom mode LEVEL 2. YOU NEED IT IN AUTO MODE SO MOST PERMISSIONS ARE AUTOMATICALLY ACCEPTED. HAVE CLAUDE AUTOMATICALLY DOING VERIFICATION, END TO END TESTING, LINTING AND THINGS LIKE THAT. This allows you to get to level 2 where you're running multiple sessions in parallel at the same time on different work trees. Most people that I see doing A GENETIC CODING THAT THIS IS PEAK. Claude writes all of nearly all of the code. "Did you read the code?" becomes "what context was the model missing and how do we solve it for next time?" Unlock: Claude proactively does work that you would have had to kick off manually before. Maintenance that used to someone now runs in the backgro step 3 to NOW THAT YOU HAVE MULTIPLE AGENTS RUNNING IN PARALLEL FOLLOWING A CERTAIN SET OF SKILLS LIKE AUTOMATIC SECURITY REVIEW AND CODE REVIEW TO MAKE THAT LOOP MORE SEAMLESS. YOU HAVE THE BEGINNING OF AN AUTOMATED CODING LOOP THAT THEN IN LEVEL 3 YOU ACTUALLY OPTIMISE AND CONTEXT ENGINEER. GET TO LEVEL YOU ACTUALLY NEED AN ADDITIONAL LOOP OUTSIDE THE CODING TO OPTIMISE THE CODING LOOP AND THINK WHAT CONTEXT COULD I HAVE PROVIDED OF THIS CODING LOOP TO MAKE IT EVEN MORE AUTONOMOUS AND HAVE MORE SEAMLESS OUTPUTS? REALLY THE PROCESS ITERATING ON THE MECHANISM OF YOUR AUTOMATED CODING LOOP TO EVENTUALLY GAIN ENOUGH TRUST IN TO LEVEL 4. The loop is fully closed and most agents are kicked off by Claude. Hundreds to thousands of agents run; you steer by intent and monitor by exception. Unlock: The super-long migration workflows check work the ri type YOU NOW TRUST YOUR CODING LOOPS THAT YOU ACTUALLY TRUST CLAUDE TO KICK OFF AUTOMATED OF OTHER CLAUDE AGENTS ON ITS OWN, SO YOU CAN RUN HUNDREDS OR THOUSANDS OF AGENTS IN PARALLEL REVIEWING ENTIRE FEATURES. INDIVIDUAL COMMITS, YOU ARE ON THE FACT THAT VIDEO RIGHT NOW PROBABLY MEANS YOU'RE ON THE RIGHT PATH, TO EXPEDITE THAT PATH. THIS STUFF IS THAT WE TALK SCHOOL COMMUNITY. IN MY BIO, TO LEARN AGENDER CODING WITH YOU.