Skip to content

The Autonomy Ladder: Which Level Is Your Workflow At?

The autonomy ladder is Dan Shapiro’s six-step scale, Level 0 to Level 5, for how much of a team’s code coding agents write and how much of it humans still read: assisted tasks, paired sessions, reviewing every diff, writing specs and checking tests, and finally a factory that turns specs into shipped software. Most developers sit at Level 2 or 3, and each guide on this site is grouped by the level it helps a team reach.

LevelNameWho writesWho readsWhat it feels like
L0By handHumanHumanEvery line is a deliberate choice
L1AssistedHuman, agent suggestsHuman, line by lineSpicy autocomplete — the human still drives
L2PairedAgent, human steers liveHuman, as it’s writtenA live pairing session, not a handoff
L3Review managerAgent, unattendedHuman, every diffA full-time reviewer, not a coder
L4Spec managerAgent, from a specHuman, tests and outcomesRequirements in, code out
L5Dark factoryAgent, end to endNo human reads the codeSpecs in, releases out

Levels after Dan Shapiro, “The Five Levels: from Spicy Autocomplete to the Dark Factory”, January 2026.

Two rungs carry warnings from the author. Level 2 is where most AI-native developers are living, and Shapiro’s line about it is that it, and every level after it, feels like you are done when you are not. Level 3 is the one almost everyone tops out at. Level 5 is rare enough to name: a handful of teams under five people, one of them identified in Simon Willison’s write-up as StrongDM’s AI division.

Answer these against the repository you ship from, not against your impression of it. The first question you answer with a no is your ceiling.

  1. Level 0 to Level 1. Did code reach disk this week that you did not type?
  2. Level 1 to Level 2. Do you hand an agent whole tasks with acceptance criteria, rather than completions inside a function you are already writing?
  3. Level 2 to Level 3. Does an agent run long enough, and often enough, that you first meet its work as a diff?
  4. Level 3 to Level 4. Is there a stop condition a machine can evaluate, so you can leave the run and check the result instead of watching it?
  5. Level 4 to Level 5. Does anything merge that no human read, and can you name the oracle that made that safe?

A level belongs to a loop, not to a company. A dependency-bump loop with a deterministic check can run at Level 4 while the feature work beside it sits at Level 2, and both answers are correct at once.

Frequently asked questions

What is the autonomy ladder?

The autonomy ladder is Dan Shapiro's six-step scale, Level 0 to Level 5, for how much of a team's code coding agents write and how much of it humans still read. It runs from every line typed by hand, through assisted tasks and paired sessions, to reviewing every diff, writing specs and checking tests, and finally a factory that turns specs into shipped software.

Which level are most developers at?

Level 2 and Level 3. Shapiro describes Level 2, paired coding, as where most AI-native developers are living, and Level 3, where the day becomes reading diffs, as the level almost everyone tops out at. His warning about both is that they feel like arrival.

What changes at Level 4?

Who reads the code, and what they read. Below Level 4 a human reads every diff. At Level 4 the developer writes the specification and the stop condition, the agent runs unattended, and the check is whether the tests passed. The reading moves from the diff to the evidence.

Is Level 5 real?

Shapiro places only a handful of very small teams at Level 5, the dark factory, where nobody reads the code; Simon Willison's write-up identifies one of them as StrongDM's AI division. It is documented, rare, and defined by the strength of its verification rather than by the absence of people.