Back in July I saw the father of Claude Code, the bald guy, post a table of AI adoption levels, from L0 to L4.
At L2, one person directs 5 to 10 agents at once, each running in its own workspace. Claude runs the tests and checks its own work before handing it over, and you only look at the final result.
L3: Claude writes almost all the code. It proactively does things you used to have to ask for, and maintenance and cleanup keep running in the background.
To reach L3, Boris lists four conditions: Claude can pull in context on its own, review speed keeps up, work is broken into routines, and Claude kicks off other Claudes.
Back then I rated myself L2.5. I was already running several agents at once and letting Claude test before handing things over, but I had only 4 automated schedules, and only half of them worked on the first try. Here’s the record from three months ago: I’m stuck at stage 2.5 of AI adoption.
Three months later I measured again, and I’d say I’m solidly at L3:
- Claude can reach my knowledge base and old conversation logs on its own
- Code it writes goes through Codex review and tests first
- Nightly batches, the morning brief, the weekly report, and error monitoring are all scheduled
- In the past two weeks, 430 Claude sessions were started by code or a schedule; only 222 were ones I opened by hand
Next up is L4. At L4, most agents are started by Claude itself, and the human only sets direction and looks at exceptions.
I’ll share how that goes when I get there.