Most engineering teams talking about AI autonomy are aiming at the wrong milestone. They're picturing Level 5 — full self-driving for code — when the goal they should actually be chasing is a trustworthy Level 3.
Borrow the self-driving ladder
Cole Medin and Dan Shapiro use the self-driving-car analogy, and it maps cleanly: from no autonomy at the bottom to full autonomy at the top. My take is simple. Most teams think they want Level 5. Most teams should be trying to earn Level 3 first.
Level 3 is the real inflection point. It's where:
- The AI does actual coding work while humans own planning and validation.
- The system stops being smarter autocomplete and becomes a production-grade tool.
- Human roles get more strategic — design, review, specification — not more mechanical.
- The productivity gains feel genuine, without asking you to trust blindly.
The "dark factory" fantasy
The dream is specs flowing in one side and finished code coming out the other, untouched by human hands. It's seductive, and it skips all the scaffolding that makes autonomy safe:
- Agent rules and conventions
- Reusable workflows and specialized sub-agents
- Separate review contexts
- Active PR review throughout the workflow
- Deterministic checks
- Real research-to-validation loops
Skip that scaffolding and here's what you get: autonomy without trust just scales mistakes. More autonomy on top of an untrustworthy system doesn't give you more good code faster. It gives you more bad code faster.
Ask a better question
Stop asking "how do we get to Level 5?" Start asking: what has to be true before we take another human out of the loop? Can you trust the specification? The review systems? The rules and boundaries? Can the system catch its own errors before they ship?
The encouraging part is that Level 3 doesn't require waiting for a better model. It's achievable today, with better planning, better review, better validation, and clearer boundaries. Earn the trust first. The autonomy is only worth having once the system underneath it deserves it.