loading…
The difference is not whether AI writes code—it is how intent becomes verified software
AI-assisted implementation exists on a spectrum. At one end, a developer describes a desired result, accepts generated code quickly, and checks it by interacting with the product. At the other, AI works inside specifications, tests, security boundaries, and review gates. Both modes can be useful. The mistake is treating prototype-level verification as production evidence.
These are not identities or opposing ideologies. A team may vibe-code a throwaway prototype to learn what users want, then convert the validated idea into an agentic engineering process.
The core difference is not prompt sophistication. It is whether success is measured through repeatable evidence.
| Artifact | How to verify it |
|---|---|
| Deterministic business logic | Unit, integration, and property-based tests |
| Generated structured output | Schema validation and contract tests |
| RAG answer | Retrieval tests, groundedness checks, citation validation, human review |
| Agent behavior | Task success, tool arguments, trajectory checks, stop-condition tests, cost limits |
| Security controls | Permission tests, adversarial cases, sandbox and policy validation |
When implementation is cheap, ambiguity becomes expensive. An agent can generate thousands of lines in the wrong direction faster than a human can review them. A useful specification therefore describes:
Chapter 5 will turn this into a complete spec-driven production workflow. Here, the foundational point is simple: AI increases implementation speed, so clear intent and verification must scale with it.
A final result can appear correct while the process was unsafe. An agent might disable a test, bypass an authorization check, or use an unapproved dependency to make the visible output pass. Reviewing only the final screen misses these failures.
Trajectory evidence can include:
Trajectory review does not mean reading every hidden reasoning token. It means observing the externally meaningful actions and state transitions the system performed.
Vibe coding has low startup cost, but repeated trial and error can create prompt patches, duplicated logic, unclear ownership, untested generated code, and fragile dependencies. This is prototype debt: decisions optimized for discovery become liabilities when the prototype quietly turns into production.
A prototype is a sketch used to decide what building to construct. Agentic engineering is the structural plan, inspection process, and controlled construction site required to let people safely live inside it.