Levels of Autonomy (Shapiro)¶
Dan Shapiro's six-stage ladder for software-creation autonomy, published Jan/Feb 2026. Zakariasson uses it as the scaffolding for the software-factory talk; andrej-karpathy has also cited it when describing Cursor's evolution.
The ladder¶
- Spicy autocomplete — tab completion, line-level suggestions. Where Cursor started in 2022–23.
- (intermediate) — multi-line / block completions.
- Pair programmer — chat loop, agent asks/answers, back-and-forth. "Where most people adopting AI tools are right now."
- Developer level 3 — AI generates the majority of code, human reviews in-loop, follows traces.
- Developer level 4 — human delegates as much as possible; reviews agent output more than code. "This is where I'm at for most software projects."
- Software factory / dark factory — black-box production. Agents plan, ship, test, build autonomously. Human provides intent and instructions only.
Why it matters¶
Locates every team in the same one-dimensional space so conversations about "where we are with AI tooling" stop talking past each other. Maps cleanly onto the org-chart analogy: L2 = pair-programming peer, L4 = individual manager, L6 = executive setting strategy.
Limits¶
One-dimensional. Doesn't capture: - Domain variance (backend can reach L5; UX-heavy work stalls at L3 because of the verification gap — see software-factory "verifiable systems" component). - Cost per unit of autonomy — L6 requires very heavy scaffolding investment up front. - jagged-intelligence — within any single "level," model capability is peaky.
Cross-references¶
- software-factory — the L6 destination
- agentic-engineering — parallel Karpathy framing
- agent-as-junior-engineer — Hashimoto's orthogonal framing (not level of autonomy but level of trust)