A product page can say “agentic AI platform” without a single agent behind it. Gartner put a number on that in June 2025: of the thousands of vendors marketing agentic AI, it estimated only around 130 had real agentic capability, and it gave the practice a name, agent washing.
That gap is the reason this vocabulary is hard to learn. AI agent, agentic, RAG, MCP, AI-native, DX, AX. They arrive in the same paragraph and sound like the same kind of word, but they are not. Some of them break the sentence if you remove them. Others survive deletion with the meaning intact.
Words you hear out loud in a meeting are a separate list, and I wrote that one already. POC, MVP, PRD and the rest live there. This post is about the words you read: press releases, slide decks, vendor pages, YouTube thumbnails.
Why can’t you learn these as one list?
Because the amount of substance behind them differs. Sort them into three layers and the confusion mostly goes away.
The technical layer is LLM, tokens, context window, RAG, fine-tuning, MCP. Delete one of these and the sentence stops meaning anything. Each points at a specific thing you can go inspect.
The working-method layer is AI agent, agentic, copilot, vibe coding. These have meanings, but the boundaries move from company to company. Two vendors will use the same word at different sizes.
The slogan layer is AI-native, DX, AX. Here is a test that works on them. Take the word out of the sentence and read what is left. “We will transform manufacturing through AX” becomes “We will transform manufacturing.” Almost nothing is lost, and you still do not know what gets changed into what.
Run the same test on a technical word and the difference shows immediately. “We connected our internal docs with RAG” collapses without RAG. A word you cannot delete carries information. A word you can delete is packaging.
What is an AI agent, exactly?
It comes down to who chooses the path. If a person laid the route out in code ahead of time, that is automation. If the model picks its own route as it goes, that is an agent.
The shortest published version of that line is in Anthropic’s Building effective agents, from December 2024. Workflows are “systems where LLMs and tools are orchestrated through predefined code paths.” Agents are “systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.” One sentence apart on the page, a budget apart in a meeting.
There is an example sitting on this site. The wheel in the sidebar pulls the eight most recent posts and picks one at random. The dice widget in one of the posts rolls a random number. Both follow a fixed rule every time. Nothing there deserves the word AI.
Put either one in a company deck, though, and it becomes “AI-powered content recommendation.” We all read that sentence weekly. Worth keeping in mind that the same Anthropic piece argues for restraint: find the simplest solution first, add complexity only when you need it, and note that this “might mean not building agentic systems at all.”
Why is “agentic” attached to everything?
Because it sells. The June 2025 Gartner prediction quoted above went further than agent washing: it expects more than 40% of agentic AI projects to be canceled by the end of 2027, blaming rising costs, unclear value and weak risk controls. I could not open the Gartner page directly, so those figures come from MarTech’s write-up of the announcement rather than the original.
None of which means agents are fake. It means there is one question worth asking when a page says agentic. Does the thing decide what to do next, or does it walk through steps somebody already wrote?
Do you need RAG or fine-tuning?
Almost always RAG. The two get compared as if they were alternatives, and they solve different problems.
RAG (retrieval-augmented generation) leaves the model alone and fetches relevant material at answer time, dropping it into the prompt. If a company says it “connected the employee handbook,” this is usually what happened. When the handbook changes, you swap the document, not the model.
Fine-tuning retrains the model itself on your data so its behaviour shifts. It suits a fixed voice, a fixed output format, or the vocabulary of a narrow field. Using it to keep facts current is expensive and goes stale fast.
So when you hear “we trained it on our own data,” ask which one they mean. Nine times out of ten nothing was trained. Something was retrieved.
What does MCP actually plug into?
It is a standard for connecting a model to outside tools and data. The official docs supply the analogy themselves: the MCP introduction says to “think of MCP like a USB-C port for AI applications.”
Why it turned up everywhere at once makes sense on a timeline. Anthropic released it on 25 November 2024, then let go of it about a year later. On 9 December 2025 the protocol was donated to the Agentic AI Foundation, a Linux Foundation fund co-founded by Anthropic, Block and OpenAI. Competitors handing the same spec to a neutral home usually means the format fight is over.
A2A (Agent2Agent) shows up next to it and does something different. Google published it in April 2025 and gave it to the Linux Foundation that June. MCP connects a model to tools. A2A lets one agent hand work to another. Different layer of the stack.
The label already sells software. While writing up Manyfast, an AI planning tool, I noticed the headline feature of its paid tier was MCP integration. As of July 2026, “supports MCP” is a normal line item on a tool’s pricing page.
Why does a long chat forget the beginning?
The context window filled up. A model reads in tokens rather than letters, and only a fixed number of tokens fit at once. Once the window is full, the oldest part falls out.
A token is a fragment smaller than a word. Languages differ here: the same content costs more tokens in Korean or Japanese than in English, so those conversations hit the ceiling sooner. That is also why a chat starts ignoring a rule you set twenty messages ago.
Hallucination is a different failure. The model does not say it does not know. It fills the gap with a plausible sentence instead, which is why names, dates, figures and links need checking every time. I said the same thing in the post on using ChatGPT day to day. Verification is a step, not an optional extra.
Three more technical words finish the layer. Multimodal means the model handles images, audio and video alongside text. On-device means it runs on the phone or laptop instead of a server, which helps latency and privacy. SLM means small language model, the trimmed-down kind that fits on such hardware. In Korea you will see the same idea written sLLM, for “smaller large language model”, which is odd enough in English that the term never crossed over.
What are DX and AX, and why have you never seen AX?
DX is digital transformation and AX is AI transformation, and the second one barely exists outside Korea. English-language tech press spells out “AI transformation” when it uses the idea at all. The two-letter habit arrived with DX, which is Japanese corporate vocabulary, and Korea carried the pattern over to AI.
The X has an explanation. Trans- comes from the Latin for “across,” which gets abbreviated to X, so transformation becomes X. DX itself is old: the concept is usually traced to a 2004 paper, which makes it twenty-two years old in 2026, not a new coinage. Japan’s Ministry of Economy, Trade and Industry made it standard corporate vocabulary there with a 2018 report warning that companies clinging to ageing legacy systems would hit a wall in 2025. That year has since come and gone.
AX is younger and it moves real money. SK C&C, a Korean IT services company, announced in May 2025 that it would rename itself SK AX, effective 1 June that year. A company changed its name to a buzzword. On the government side, the Ministry of Trade, Industry and Resources runs a manufacturing AI programme called M.AX, whose alliance grew from 1,000 member organisations at its September 2025 launch to 1,300 three months later, with 700 billion won (roughly 500 million US dollars) budgeted for 2026.
Korea also moved some of this vocabulary into statute. Its AI framework act took effect on 22 January 2026, and “generative AI” and “high-impact AI” are defined legal terms under it. Slogan one year, legal category the next.
The 18 terms in one table
Grouped by layer. The last column is what happens to the sentence when you delete the word.
| Term | Layer | What it means | Delete it and |
|---|---|---|---|
| LLM | technical | Large language model, the engine behind most of this | sentence breaks |
| SLM | technical | Small model, fewer parameters (sLLM in Korea) | sentence breaks |
| Token / context window | technical | The unit it reads, and how much fits at once | sentence breaks |
| Hallucination | technical | Confident invention where knowledge is missing | sentence breaks |
| Multimodal | technical | Images, audio and video, not just text | sentence breaks |
| RAG | technical | Fetches source material at answer time | sentence breaks |
| Fine-tuning | technical | Retrains the model itself on your data | sentence breaks |
| MCP | technical | Standard connecting a model to outside tools | sentence breaks |
| A2A | technical | Standard for one agent handing work to another | sentence breaks |
| On-device | technical | Runs on your hardware, not a server | sentence breaks |
| AI agent | method | Picks its own path to finish a task | size varies by vendor |
| Agentic | method | The adjective for that behaviour | often just a label |
| Copilot | method | Assists rather than replaces (also a product name) | depends who says it |
| Vibe coding | method | Describing what you want and letting AI write it | quality bar varies |
| Human in the loop | method | A person checks before the next step runs | where you cut matters |
| DX | slogan | Digital transformation | meaning survives |
| AX | slogan | AI transformation | meaning survives |
| AI-native | slogan | Built assuming AI from day one | meaning survives |
Vibe coding sits awkwardly between layers, so I filed it under method. It was coined in February 2025 and became Collins Dictionary’s word of the year that November, defined as “the use of artificial intelligence prompted by natural language to write computer code.” Nine months from a post to a dictionary entry.
How I read them now
Slogan words get deleted, technical words get looked up, method words get a question back to whoever said them. Knowing that the three need different handling is most of the benefit.
This is not a case for cynicism. Labels move money. A 700 billion won programme got the name M.AX, and a listed company put AX on its front door, and because of those names budgets and teams actually moved. The mistake is expecting the sentence itself to tell you something.
I have been caught by the gap between a label and the thing it names. Testing tools for making your own font, I hit one whose FAQ said its output was OFL 1.1, and the licence strings inside the file I downloaded turned out to be GPL v3 alongside OFL. The only way I found out was opening the file. Same rule applies here. Read the label, then go look at the thing.