The Agentic Engineering Operating Level Framework

The Five Levels of Agentic Operation

IndyDevDan introduces a framework for mapping where engineers and their agents focus their attention. The hierarchy spans five levels:

1. Code Primitives: Lines, blocks, functions, and types.
2. Code Structure: Files, modules, and directory organization.
3. Data and Execution: Database schemas, tables, and script-based control.
4. Delivery and Intent: Planning, documentation, and application-level behavior.
5. Agentic Systems: AI developer workflows and software factories.

The Tradeoff: Leverage vs. Control

The core insight is that moving up the stack provides leverage and speed, while moving down provides control and understanding. The "trap" is attempting to scale a system via high-level agents without understanding the lower-level primitives. Effective engineers do not park at one level; they possess the range to move up and down based on the specific problem's requirements.

When to Choose Leverage vs. Control

Engineers should opt for leverage when the domain is familiar, the work is repetitive (three instances make a pattern), and evidence supports automation. Conversely, they should drop down to control when domain expertise is thin, the system is unfamiliar, risk is high, or the task is "out of distribution" for the model. The author argues that true expertise is what expands the model's distribution, allowing the engineer to guide the agent through territory it would otherwise fail to navigate.

The Fallacy of Vibe Coding

Many engineers oscillate between two extremes: "vibe coding" (only inspecting the final application output) and "loop engineering" (over-relying on automated agent loops). The author suggests that the most capable engineers treat agentic engineering as a distinct craft, mastering prompt, context, and harness engineering to build systems that they can then effectively manage from the top down.