The vocabulary of AI changes every six months. Here is the map.
Prompt engineering, context engineering, agentic, harness, loop, graph. Nine names in six years for roughly the same thing. Who coined what, on which date, and what actually matters when you are not an engineer.
Between 2020 and today, the central discipline of working with AI has been renamed nine times. Not improved nine times: renamed. The same mechanic repeats every time. Someone posts a sentence, the word catches on, a wave of articles explains it, courses get sold, and six months later another word replaces it.
The problem is not that these words are empty. Most of them describe a real shift. The problem is that nobody dates anything, and the same attribution errors travel from article to article, copied without checking.
So here is the map. Every name, its date, what it adds in one sentence, and above all who coined it, with a link to the original source.
Discipline
Prompt engineering
Word the instruction so the model's output becomes usable.
Claimed by four different lineages
Discipline
AI engineering
The LLM becomes a product component called by code, not a chat toy.
Named by Shawn “swyx” Wang
Discipline
Flow engineering
Split the task into checkable steps instead of hoping for a perfect one-shot.
Named by Ridnik, Kredo and Friedman
Practice
Vibe coding
Accept the output without reading the diff, to move very fast.
Named by Andrej Karpathy
Discipline
Context engineering
Provide everything needed for the task to be plausibly solvable.
Named by Tobi Lütke, amplified by Karpathy
Methodology
Spec-driven development
The versioned spec becomes the source of truth, not the chat history.
Named by GitHub, with Spec Kit
The turn
Karpathy: I have never felt this much behind. The profession is being refactored.
Said by Andrej Karpathy
Discipline
Agentic engineering
You no longer write the code 99% of the time, you orchestrate agents.
Named by Andrej Karpathy
Discipline
Harness engineering
Not what the agent sees, but what the system prevents, measures and repairs.
Named by Ryan Lopopolo, or Mitchell Hashimoto
Discipline
Loop engineering
Stop being the one who prompts the agent, become the one who designs the system that prompts it.
Named by Peter Steinberger, formalised by Addy Osmani
Emerging term, not settled
July 2026 Six weeks after loop
Claimed discipline
Graph engineering
Design the topology wiring several loops together, not the execution of one. Three competing meanings in 48 hours, and LangChain claims three years of prior art.
Attributed to Josh Simmons, with no independent confirmation
How to read this timeline
The name at the bottom of each card is clickable. It points to the original source: the research paper, the blog post, the tweet. That is the only way to check rather than take my word for it.
Words with a dotted underline carry a tooltip. Hover or tap for the definition, without leaving the page.
The first four optimise the input, the last four optimise the environment
This is the most useful reading of the whole timeline. Prompt engineering, RAG, flow engineering, context engineering: all of them try to phrase what we hand the model better. Agentic, harness, loop, graph engineering: all of them try to build what surrounds the model so it can work alone without breaking things.
The turn happened in December 2025. Karpathy, who had named vibe coding a year earlier, posted that he had never felt this far behind, and that the profession was being refactored. The following six weeks produced two new names.
The two terms of summer 2026
Loop engineering, in June. The idea fits in one sentence: you stop being the person who talks to the AI, you become the person who builds the system that talks to it. The head of Claude Code at Anthropic puts it bluntly: he no longer prompts, he writes loops that prompt on his behalf.
Graph engineering, in July. Six weeks later. And this is where it gets interesting.
What this changes for you, concretely
If you are not an engineer, the timeline comes down to three decisions.
Do not chase the latest word. Loop engineering lasted six weeks before its death was announced, and the announcement was a joke. A course sold on the term of the month mostly sells the term of the month.
Look at the layer, not the name. The real question is not “am I doing context engineering”. It is: does my problem come from what I hand the AI, or from what surrounds the AI? The timeline tells you which half you are in.
Check the dates before you quote them. If you write about these subjects, go back to the source. Half the articles explaining harness engineering credit the term to someone whose original post nobody has been able to find.
This page gets updated
It will be revised every time a term is added or a status changes. The date at the top of the article marks the last analysis. Unsettled entries stay in the dark box at the end until they are confirmed.
If you spot an error, or if you have found one of the two sources marked with a dotted underline, tell me. That is exactly the kind of correction that makes this page useful rather than redundant.