The spec is the contract

In a Cognitive Enterprise, the division of labor shifts: humans write specs, agents write implementation.

That only works if the spec is real. A spec is: a named input, an expected output, measurable acceptance criteria, and what escalates to a human. Without acceptance criteria, what you have is a wish, not a specification.

The payoff shows up in maintenance. When someone wants to change the system’s behavior, they change the spec and let the rest derive from it. In most projects, the opposite happens: someone hunts for an adjective in the prompt, tweaks a sentence, and finds out three weeks later that something else broke.

Before the spec comes research, a declared discovery phase that ends with a short findings document. It may look like bureaucracy, but it is what prevents building an elegant solution to the wrong problem. If the research did not change any assumptions, either the problem was trivial or the research was shallow.