A prompt was AI’s proof of concept. It showed that a large language model could be useful. Every stage since has addressed a weakness exposed by the stage before it.
Nine stages of development
1. Prompts
Natural-language questions demonstrated that general-purpose language models could produce useful explanations, drafts and comparisons.
2. Prompt Engineering
Structured instructions, examples and context improved relevance and consistency.
3. Prompt Chaining
Connected prompts divided complex work into manageable stages.
4. Loop Engineering
Iteration, critique and revision enabled systems to refine outputs rather than stopping after one answer.
5. Agent Engineering
Specialised AI workers were given roles, tools and objectives.
6. Graph Engineering
Agents, tasks, decisions and dependencies were organised into structured workflows.
7. Evidence Engineering
Attention shifted from fluent outputs to claims, evidence, counter-evidence and uncertainty.
8. Governance Engineering
Permissions, policies, approvals, boundaries and audit trails were engineered into the operating system.
9. Reliance Engineering
The final question became whether the result was fit for its intended purpose and safe to act upon.
The lesson
Capability evolution and control evolution must progress together. A more capable system without stronger verification can simply create larger, faster and less visible failures.
Revision history
Version 1.0 — 26 July 2026: Initial publication.