The five levels of AI-assisted software development are a model proposed by Goetz Markgraf, a consultant at codecentric, for describing how far AI takes part in programming, drawn as a parallel to the levels of autonomous driving. He first presented it in a talk at a conference on AI-assisted coding and described it in a February 2026 blog post, arguing that without such a classification people talk past one another about “AI in programming”. The five levels are AI code assistance (level 1), AI code generation (level 2), agentic AI development (level 3), spec-driven development (level 4) and autonomous product evolution (level 5).
Usage
In Markgraf’s account:
- Level 1, AI code assistance — web-based AI chats such as ChatGPT, Claude or Gemini. The developer asks a question, copies the answer into their own code by hand, and builds the context themselves by copy-and-paste or uploading documents; the AI acts as a kind of “StackOverflow on demand”. His driving parallel is a simple cruise control.
- Level 2, AI code generation — IDE plugins such as GitHub Copilot, JetBrains AI Assistance or Continue.dev propose completions, offer chat, and can generate whole code blocks from an instruction; the developer critically reviews each proposal. The parallel is adaptive cruise control.
- Level 3, agentic AI development — an agent such as Cursor, Windsurf, Claude Code or OpenCode takes on a task from a natural-language instruction and works through several steps on its own: finding and reading files, writing code, possibly running tests and fixing errors. The developer supervises, intervenes and checks the result. The parallel is lane-keeping with automatic lane changes on the motorway.
- Level 4, spec-driven development — human and AI first work together to understand the task and write a detailed plan (a spec), after which the AI generates fairly complex code or changes existing code largely on its own. It uses the same agents as level 3 and differs only in how they are used; the final result must still be reviewed intensively and tested by hand. The parallel is autonomous driving in defined areas, such as a parking-garage service or motorway traffic jams. See Spec-driven development.
- Level 5, autonomous product evolution — a whole product is written and developed further exclusively by AI, including operation and monitoring, and nobody looks at the code. He calls it still largely hypothetical: tools such as v0, Bolt, Lovable, Emergent or Replit generate whole applications from descriptions, but in his view do not yet produce production-ready solutions or run them on their own. The parallel is full driving automation with no steering wheel.
Markgraf says the model carries no ranking — level 3 is not automatically better than level 2 — but that efficiency rises with each level. He reports that speakers who followed him at the conference spontaneously adopted the model to place their own topics within it.
AI readiness
The post pairs the model with the preconditions it calls AI readiness: the readiness and ability
of a team, a codebase and an organization to use AI effectively and without risk. Its examples are
clean code and a clear architecture, made known to agents at levels 3 and 4 through instruction files;
automated tests, which it calls the most important safety net, including not letting the AI switch
off or delete tests; documentation such as README files, a docs/ folder in Markdown or arc42;
team skills, including jointly maintained instruction files and a Definition of Ready or Done;
isolating agents from the host from level 4 onwards, for example with Docker containers, to limit the
damage from Prompt injection; and a paid AI subscription whose terms rule out using the
processed code for training.
When It Applies
The model is one practitioner’s orientation aid rather than an empirically derived scale: its stated purposes are helping a team establish where it stands and where it wants to go, setting realistic expectations, and giving people a shared vocabulary. Its support is the author’s own experience and his report of how it was taken up at one conference. In his account level 4 is separated from level 3 by a way of working rather than by new technology.