The autonomy ladder
Five levels, borrowed from the way the industry talks about self-driving cars. The borrowing brings a useful frame and one dangerous habit with it.
In January 2026 Dan Shapiro published a five-level scale for AI-assisted development, modelled on the NHTSA and SAE levels of driving automationprimary. It was amplified within the weekreputable and the vocabulary stuck, because it gave people a way to say where they were without claiming more than they had.
a human still reads the output only machines read the output
What each rung actually asks of you
Read the ladder as a ladder of reading, and the rungs stop being abstract.
L0 and L1 cost you nothing structurally. The AI is a better search box, or completes the line you were already typing. Your relationship to the code is unchanged: you wrote it, you know it.
L2 is pair programming, and Shapiro’s estimate is that around 90% of AI-native developers live hereprimary. You still read everything, because everything arrives in front of you as you go.
L3 is the first rung with a real cost. The AI writes most of the code and you become a full-time reviewer. Nothing about the work is unfamiliar — it is code review, which the industry has done forever — except the volume, and the fact that you did not build the mental model first. Reviewing code you would have written is fast. Reviewing code you would not have written is slow, and this is where the first complaints about verification start.
L4 moves you again: you write the specification and check the tests, which is much closer to engineering management than to programming. Shapiro places himself here. This is where most of the tooling in chapter 4 is aimed.
L5 is the dark factory. Specs in, software out, no human review. Shapiro notes that only a handful of small teams have reached it, and mid-2026 evidence has not moved that muchprimary.
Where teams actually sit
Nearly all production deployments are at L3 or L4, and that is true of the most sophisticated ones. Stripe merges over 1,300 agent-authored pull requests a week and reviews every one of themprimary. Spotify runs a fleet that merged 1,500 PRs in nine months and now merges around a thousand every ten days, all CI-verified and human-approvedreputable. At Cursor, 35% of internally merged pull requests come from cloud agents, and the company describes code review as the remaining human bottleneckprimary.
The shape is consistent: unattended execution, attended approval. The autonomy is in the doing, not in the deciding.
The habit the car metaphor smuggles in
The driving-automation scale is a good loan and a slightly dangerous one, because it carries an assumption: that the levels are a progression, and that a serious organisation is climbing them. In cars that is roughly true. The destination is the same at every level, the vehicle is one system, and the top of the ladder is unambiguously the goal.
Software is not one system. A codebase is thousands of loops with wildly different blast radii, and there is no reason for them to sit on the same rung. The nightly job that fixes one lint violation and the change to your billing logic are different jobs with different verifiability, permanently — no journey connects them. Chapter 8 makes this concrete: the skilled work happens loop by loop, deciding where each switch goes, and climbing has nothing to do with it.
Held loosely, the ladder is the best shorthand the field has. Held tightly, it turns a distribution of decisions into a single number to be maximised — and a single number to be maximised is how token utilisation became a proxy for progress in chapter 6.
Do not give your organisation a level. Give each loop one, write it down next to the loop, and re-derive it when the loop's check changes. An org-wide level is a slogan; a per-loop level is an operating decision you can defend.