Confidence: medium. Evidence: benchmark and case study. Last substantive change: 2026-07.
The factory can only build what it can be told to build. This subsystem owns how intent enters the system and how much of the human's reasoning survives the trip into a specification an agent can act on.
The conclusion
A prose prompt is not a specification. Requirements should be executable enough to test, yet rich enough to carry rationale, non-goals, risk, and uncertainty alongside the generated code. The quality of the output is bounded by the quality of the intent, so specification quality is where much of the human effort now concentrates.
How the thinking got here
Issue text gave way to detailed natural-language specs, then to structured run contracts, and then to guarantees and verification obligations carried beside the code. Each step tried to preserve more of what a human meant, not just what a human typed, because visible specs and tests turned out to be gameable and incomplete when they carried too little intent.
Credible alternatives, and when each is right
| Approach | Right when |
|---|---|
| Conversation-first clarification | ambiguous, exploratory work |
| Spec-first documents | stable, well-understood requirements |
| Examples and scenarios | behavior is easier to show than to state |
| Formal methods | small, high-criticality cores |
| Test-driven contracts | acceptance is crisply checkable |
| Reverse-specification | recovering intent from a brownfield system |
Where it fails and what we still don't know
The risk is confident construction of the wrong thing: a spec that reads well, passes its own tests, and misses what the human actually needed. No universal representation of intent has won. Open questions include spec-quality metrics, change propagation, stakeholder sign-off, ambiguity detection, requirements provenance, and whether an agent can safely propose requirements it will later implement.
What would change our mind
A measured, repeatable way to score specification quality that predicts downstream defects would turn this subsystem from craft into engineering.