Levels of AI coding autonomy: what changes from level 0 to level 5

The levels of AI coding autonomy are a scale from 0 to 5, proposed by Dan Shapiro in January 2026, that describes how much of the work of programming moves from a person to AI. At level 0, in his words, "not a character hits the disk without your approval." At level 5, which he calls the dark factory, the process becomes "a black box that turns specs into software." Ithaca operates at level 4 with a rule of its own: coding agents execute, people on the team read what goes to production, and every session is logged.

Published 10 October 2026

Who this page is for

For CTOs and engineering leaders who need to say where their team stands with AI today and how far it makes sense to go. The scale gives that conversation a shared vocabulary, including with risk and compliance.

Where the scale comes from

Dan Shapiro, CEO of Glowforge and a research fellow at Wharton, laid out the levels in "The Five Levels: from Spicy Autocomplete to the Dark Factory" on January 23, 2026. He had seen dozens of companies struggle to put AI to work writing code, and each of them "has moved through five clear tiers of automation."

He modeled the scale on the levels of driving automation, which also start at zero, so five levels make six rungs. Five days later, Simon Willison summarized the scale on his blog, and other writing on coding agents has cited it since.

The scale at a glance

The six rungs of Dan Shapiro's scale. The level names come from the post.
Level	Name in the post	Shapiro's car comparison	What the person does, in Shapiro's description
0	Spicy autocomplete	"Your parents' Volvo"	Writes all the code. AI is "a search engine on steroids" or a suggestion accepted with tab
1	Coding intern	"Lanekeeping and cruise control"	Writes the important parts and offloads discrete tasks, like a unit test or a docstring
2	Junior developer	"Autopilot on the highway"	Pairs with the AI inside an AI-native coding tool
3	Developer	"A Waymo with a safety driver"	Reviews the code the AI writes: "Your life is diffs"
4	Engineering team	"A robotaxi": while it drives, you can do something else	Writes and argues over the spec, plans and reviews plans, and comes back hours later to see if the tests pass
5	Dark factory	"Not really a car any more"	Hands in the spec. The process is "a black box that turns specs into software"

Each rung takes one more piece of the work out of human hands. From level 3 up, the question that matters most to a company is who reads the code before it goes into production.

Each level in practice

Level 0: spicy autocomplete

"Whether it's vi or Visual Studio, not a character hits the disk without your approval." AI might show up as "a search engine on steroids" or as a suggestion accepted with tab now and then, but in Shapiro's words, "the code is unmistakably yours."

Level 1: coding intern

Shapiro calls the AI "your AI intern." The person still writes the important parts and offloads specific, discrete tasks: "Write a unit test for this." "Add a docstring." The tool could be anything from copy-paste ChatGPT to Copilot. There are speedups, but "your job is unchanged. You're still moving at the rate you type."

Level 2: junior developer

The AI becomes "a junior buddy to hand off all your boring stuff to," inside an AI-native coding tool rather than a chat window. This, Shapiro writes, "is where 90% of 'AI-native' developers are living right now." He adds a warning: "level 2, and every level after it, feels like you are done. But you are not done."

Level 3: developer

"You're not a senior developer anymore; that's your AI's job. You are… a manager." The coding agent is "always running multiple tabs," and the day goes to code review: "Your life is diffs." This is where Shapiro places the human in the loop. For many people, he writes, it feels like things got worse, and "almost everyone tops out here."

Level 4: engineering team

The person becomes a PM (in a footnote, Shapiro leaves open whether that means program, project or product manager): "You write a spec. You argue with it about the spec." You craft skills, plan schedules and review plans. "Then you leave for 12 hours, and check to see if the tests pass." The post's illustration for this level reads: "The AI runs unattended for long periods, handling complex tasks. The human trusts the system's self-checks, only checking the final results much later." Shapiro places himself at this level.

Level 5: dark factory

"It's not really a car any more," Shapiro writes. "It's a black box that turns specs into software." In the illustration, "the engineer manages the goals and the system, not the code." The name comes from the Fanuc Dark Factory, "the robot factory staffed by robots," which is dark because "humans are neither needed nor welcome." He says he knows "a handful of people who are doing this," in small teams of fewer than five, and that it "will likely be our future."

From level 3 to level 4: the reading moves

At level 3, a person reviews everything, all the time. That is an old problem. In "Ironies of Automation" (1983), Lisanne Bainbridge of University College London noted, citing vigilance studies, that even a highly motivated person cannot keep effective visual attention on a source where very little happens for more than about half an hour. Her paper was about process plants and flight decks; the parallel with reviewing agent output is our reading.

At level 4, the question is no longer what a person watches but where a person decides. On Ithaca's platform, control sits at three points:

Before the code. The team writes the specification and sets the autonomy limit: what the agents may do without prior approval and what needs a person's sign-off first.
Before production. People on the team read what goes to production before it goes in, rather than watching every step as it happens.
In the session record. Every session records the request, the steps, the model, the cost, the result and the authorship. A review starts from what the agent did, and not only from the final diff.

Why Ithaca runs at level 4

Ithaca operates an agentic development platform at level 4 of the autonomy scale Dan Shapiro proposed. In Shapiro's description, level 4 is the spec rung. The text of his post does not say who reads the code at that level, and its illustration has the human trusting the system's self-checks and checking the final results much later. Ithaca's practice adds a rule: coding agents execute, people on the team read what goes to production, and every session is logged, with its authorship. Level 4, the dark factory and the autonomy scale itself have short definitions in the glossary of software development with coding agents.

The reasons are practical:

Someone has to answer for the code. In a large company, the risk committee asks who decides what the agents do, what each session leaves on record, and who reads the code before it reaches production. Until all three have an answer, the conversation about speed goes nowhere. The checklist is in AI governance in software development.
Code that people read stays maintainable. When something breaks at three in the morning, the damage depends on how long it takes to find the cause. What each session records is covered in Observability for coding agents.
The execution gain stays. Coding agents take on the execution that does not need a person; the team spends its time specifying, reviewing and deciding.

The limit is not fixed. The team decides what the platform does on its own at each stage and keeps adjusting that limit after the deployment. The rules and limits the agents work under are a discipline of their own, described in Harness engineering for enterprise coding agents.

Why level 5 is not the goal

Level 5 exists, and some teams work that way. Simon Willison, who talked to one of them, listed its key traits: "Nobody reviews AI-produced code, ever. They don't even look at it." Much of the agents' work goes into testing, tooling and simulating the systems the software depends on, and the people design that system instead of reading the code.

For Ithaca, level 5 is not the goal, for three reasons.

When it breaks, someone has to understand the code. In July 2025, Dex Horthy had the model write the code with nobody reviewing it. Four months later, the team shut it down and threw the whole system out. Production broke, the model could not find the root cause, and it took days of digging to find a primary key routed wrongly through the whole codebase, plus three weeks to re-onboard to a codebase no one had ever read.
People who only watch lose the skill. Bainbridge pointed out that physical skills deteriorate when they are not used, and that "a formerly experienced operator who has been monitoring an automated process may now be an inexperienced one." In a system nobody reads, nobody is ready to take over. Applying that to software is, again, our reading.
A large company needs an identifiable author. Systems with years of business rules, regulated data and audits need, for every change, a record of who asked, who supervised and who approved the move to production.

None of this is a prediction about level 5. It is the position the platform operates from today. The full argument is in What a dark factory is, and why the platform is not one.

How to tell which level your team is at

Three questions settle most cases:

Who writes most of the code that reaches production? If people do, the team is somewhere between levels 0 and 2. If AI does, it is at level 3 or above.
What does the person spend the day on? Reviewing diffs is level 3. Writing and arguing over specs and checking results is level 4.
Does anyone read what goes to production? If the answer is no, by design, that is level 5.

A fourth question is not on the scale, but it decides whether a team can move up a rung and stay in control: what does each session leave on record?

Dr. Consulta on the platform

Dr. Consulta modernized its subscription platform, a system with eight years in production, using the agentic development platform, with its own team supervising. Compared with the period before the modernization, infrastructure cost is 75% lower, and one representative feature went from about two weeks to up to three days. That feature is an example, not a measurement of the whole backlog. Read the Dr. Consulta case study

Frequently asked questions

What are the levels of AI coding autonomy?

The levels of AI coding autonomy are six rungs, from 0 to 5, that describe how much of the work of programming moves from a person to AI. Dan Shapiro proposed the scale in January 2026, modeled on the levels of driving automation. It runs from level 0, where nothing is written without the programmer's approval, to level 5, the dark factory, where the process is "a black box that turns specs into software."

What level does Ithaca operate at?

Level 4. Coding agents execute; the team writes the specification, sets the autonomy limit and reads what goes to production. Every session is logged with the request, the steps, the model, the cost, the result and the authorship. Reading before production is Ithaca's rule, not part of Shapiro's definition.

What is the difference between level 3 and level 4?

In Shapiro's description, at level 3 the person spends the day reviewing the code the AI writes, and at level 4 the person writes and argues over the spec and comes back later to see if the tests pass. On Ithaca's platform, people on the team still read what goes to production at level 4, with the session record in hand.

What is level 5, the dark factory?

On Shapiro's scale, it is the rung where the process becomes a black box that turns specs into software, and nobody reads the code. He says he knows a handful of people working this way, in teams of fewer than five. Ithaca's platform stops before it. More on the dark factory page.

Is the scale an official standard?

No. It is a proposal Dan Shapiro made in a blog post. It is not a standard or a certification. It helps frame the conversation; the autonomy limit for each system is still the team's decision.

Doesn't human supervision cancel out the speed gain?

It doesn't have to. Coding agents take on the execution, and the team keeps the specification, the review and the decisions. At Dr. Consulta, compared with the period before the modernization, one representative feature went from about two weeks to up to three days.

Sources

Dan Shapiro, "The Five Levels: from Spicy Autocomplete to the Dark Factory," January 23, 2026The scale, the level names, the car comparisons and the description of each level.
Simon Willison, "The Five Levels: from Spicy Autocomplete to the Dark Factory," January 28, 2026A summary of the scale, and the traits of a team working at level 5.
Gergely Orosz, "Context engineering with Dex Horthy," The Pragmatic Engineer, July 15, 2026The four months, the shutdown and the three weeks of re-onboarding.
Lisanne Bainbridge, "Ironies of Automation," Automatica 19(6), 1983Vigilance, and the loss of skill in people who only monitor.