03 — Terms mapped to the continuum
The continuum has five bands, from vibe coding to governed multi-agent work. The literature uses many other names for AI-assisted software development. This page shows which band each one belongs to. The second list covers practices that apply in every band.
AI agents plan, edit, run checks, debug and iterate toward a goal using tools. Humans set the goals and review.
The human stays the primary author and driver; the AI acts as a continuous navigator, offering suggestions, explanations and issue spotting.
Specialized agents (planner, coder, tester, reviewer) collaborate under an orchestration layer.
The red-green-refactor cycle with AI generating tests, edge cases, minimal implementations and refactors. Humans still define correctness through the tests.
The system retrieves relevant files, APIs and docs from the real codebase before generating or editing.
You provide input–output examples; the system synthesizes a program that matches them.
Structured, sequential prompting in which the human still breaks the work down. This is distinct from giving in to the vibes.
Managing what goes into the context window (files, specs, constraints, prior decisions) so the model stays grounded. It becomes critical once you leave pure vibe mode.
Versioned, machine-readable specs used as the source of truth from which code, tests and docs are derived or checked.
Vibe-code to explore, then formalize the chosen direction into a living spec before production work. Most experienced practitioners recommend this over choosing one approach exclusively.
Linking requirements → stories → specs → code → tests so decisions can be audited or reversed. Essential for regulated, multi-person or long-lived systems.
Automated reviewers, contract testing, checks that spec and code agree, and security scanning of AI-generated output. AI review can inherit the code’s blind spots, so it supplements human review rather than replacing it.
Scoped credentials, sandboxes and explicit approval points (before merge, before production data, before anything reaches students), placed where the risk is.
Working code that nobody on the team understands is a liability that grows over time. In teaching, that missing understanding is the learning failure.
Recording what was AI-generated, with which model and under what instructions, for academic integrity, reproducibility and license review.
A useful grouping. The exploration and speed end holds vibe coding, pure prompt-driven work and light AI assistance. The middle holds agentic coding with review, AI pair programming and AI-augmented TDD. The production and durability end holds spec-driven development, multi-agent systems with strong verification, and governance-heavy workflows.