Human-AI Autonomy Scale
AI Development Autonomy Scale
How much engineering execution is delegated to AI, and what responsibility remains human.
From AI in the void to Human in the void.
AI in the void · Human baseline
00 / Human-led
You do the development work without generative AI.
AI's role
None. Compilers, linters, CI and ordinary IDE tools still count as human-led work.
Human's role
Choose, write, check and ship the changes.
Loop: real-time · local · IDE
For example: You implement a feature, write tests and review the diff yourself, without generative AI.
Understanding the scale
A level describes an approximation of a typical workflow, not a person or a model. A real case scenario may span multiple levels: a level 7 initial draft being refined with AI at level 5 or lower. A level's perception also depends on the frequency of AI involvement and human authored contributions.
Human in the void (not needing a human) means just routine engineering execution. It does not extend to QA, release acceptance or product ownership.
Loop duration illustrates instruction and feedback intervals at levels 1-8, and oversight or acceptance intervals at levels 9-12. It is not a measure of AI processing speed.
2022-2026 / observed workflows
How this sample changed
Mean inferred minimum level of linked, self-selected posts. Not the average developer or a population trend.
2026 / 25 accounts
4.96 / 12
Mean observed floor, not maximum capability.
Typical practice in this sample
Parallel agents and background bug fixes coexist with advice-only use and manual review.
Models actually named
Opus 4.6, GPT-5.4, DeepSeek v4 Pro, Astra and Luna named; most unspecified.
Independent coding evaluation
SWE-bench Verified leaderboard (2026): Feb 2026, mini-SWE-agent v2.0: Opus 4.5 resolved 76.8%; Gemini 3 Flash and MiniMax M2.5 each 75.8%.
External comparisons, not results for these accounts. Tasks and setups differ across years; live leaderboards may change. Scores are not autonomy levels.
25 dated source posts for 2026
About the Scale
Most frameworks ask: “What can the AI do?” This one asks: “Who is doing the work — the human or the AI?”
The AI Development Autonomy Scale is the development edition of the Human-AI Autonomy Scale, a shared reference for developers, teams and job descriptions. It measures delegated engineering execution and the engineering responsibility that remains human.
It describes how work is performed, not model capability or ownership of product strategy. Human QA, technical oversight and product acceptance can remain even when AI owns engineering execution.
In this edition, “human” refers to the developer executing the delegated engineering work. “Human in the void” means that work no longer requires a developer to perform or supervise it, not that all human oversight disappears. QA, technical leadership, product ownership and release acceptance may still remain. At level 9, developers delegate routine execution; at level 12, the engineering function no longer requires human developers.
Same tools. Completely different reality.
Two teams can use the same assistants and still operate at very different autonomy levels.
- Hiring clarity: “This role operates at level 6–7.”
- Team alignment: “We want to move from level 4 to 6.”
- Expectation setting: avoid “we use AI” mismatch.
- Strategic planning: map transition paths between levels.
What's level 13?
This development edition is the first part of the intended Human-AI Autonomy Scale, covering the whole organizational structure of software companies. Future editions will explore how responsibility shifts between humans and AI across product, business functions and company leadership, beyond engineering execution.