In agentic engineering, code is the agent’s output but the specification is the human’s. A specification is a precise document that gives the agent enough information to make autonomous decisions correctly, so that the human does not need to intervene at every step. A feature list is something different.

The formal benchmark behind a complete specification is the five-criteria definition of an algorithm Donald Knuth set out in The Art of Computer Programming: finiteness, definiteness, input, output, effectiveness. A natural-language briefing cannot reach the precision of a formal algorithm. The trade-off in agentic engineering is to give up some formal precision in exchange for the ability to describe higher-level intent. A briefing that ignores these properties is a wish, not a specification.

A complete specification contains the goal and its business rationale, success criteria that are verifiable and not aspirational, architectural guidance that gives the agent strategic direction without dictating implementation steps, a validation approach that defines how correctness will be checked, and known pitfalls that an experienced developer would warn a colleague about.

Categories of specification gap map to categories of agent failure. Missing success criteria produce agents that terminate when the code runs rather than when it works. Missing business rationale produces agents that solve the literal problem in ways that violate the intent. Missing architectural guidance produces agents that choose structurally correct but contextually wrong approaches. The reverse-interview methodology is the practical technique: prompting the agent to ask questions until it has enough information to write the specification itself.