A field guide to AI buzzwords
For each popular term: what it actually names, and how much of it is marketing.
Words on this page (7)
- vibe coding
- : Building software by prompting and accepting results with little review.
- agent
- : A model run in a loop that chooses its own tools and steps until the task is done.
- tool
- : A function the model can ask your program to run, described by a name, description and argument schema.
- AGI
- : Artificial general intelligence: a loosely defined label for AI as capable as people across most tasks. There is no agreed test.
- context engineering
- : Deciding what goes into the model's context window, and when.
- token
- : A piece of text the model reads and writes, usually a word or part of a word. About 4 characters of English on average.
- skill
- : A folder of instructions and scripts an agent loads only when a task needs it.
Vibe coding
Real: building by prompting and accepting output with little review, fine for prototypes. Limit: without verification you cannot trust it for anything that matters. engineering is the same with specs, tests and review.
Agentic
Real if it means the model chooses its steps. Hype if it is stuck on a fixed chatbot flow. Ask what decisions the model makes at runtime.
AGI
A loose label with no agreed definition or test. Useful as a direction, useless as a product requirement. Treat claims about its date as opinions.
Context engineering
Real: deciding what enters the window and when. It is the most concrete of the newer terms because you can measure and results.
Multi-agent swarm
Real in narrow cases: parallel research, large audits and independent verification. Often overkill. One agent with good tools is simpler and cheaper, and every extra agent adds coordination cost.
Skills
Real: folders of instructions and scripts loaded when relevant. A packaging convention, not a new capability.
Memory
Real: storing facts outside the window and retrieving them. Mostly files plus search. The hard part is deciding what to keep and when it is stale.
Reasoning
Real: models trained to spend tokens on thinking before answering. It improves multi-step problems. It does not mean the model understands in a human sense.
"10x" and "replaces developers"
Studies are mixed. One 2025 randomized study by METR followed 16 experienced open-source developers on 246 real issues and found tasks took 19 percent longer with the AI tools of early 2025, partly from the time spent verifying. It is a small sample and the tools have since changed, so treat it as one data point. Gains are real on well-specified, verifiable work and smaller elsewhere.
A worked example: translating a pitch
Pitch: "Our multi-agent swarm autonomously builds your product." Translation: several model calls coordinated by code. Questions: what does each agent do, what checks the output, what is the pass rate, what does a run cost, what happens when it is wrong? If the answers are vague, it is a demo.
A quick glossary of caution
Real mechanism with a hype halo: agents, reasoning models, , , memory. Mostly packaging: "swarm", "autonomous employee". Undefined: .
A second example: "our assistant has memory"
Mechanism, with an invented example:
- After a chat, a step asks the model which facts are worth keeping, such as "user prefers metric units".
- Those facts are saved to a file or database.
- In the next chat, a search finds the notes that match the new message and pastes them into the window.
That is memory: a notebook the system reads before it answers. The failure is a stale note. You tell the assistant you moved from London to Paris, but the old note stays, so it keeps suggesting umbrellas for London. To confirm, change a fact and ask a question that depends on it in a new chat.
When the buzzword is the wrong tool
A task has three steps: fetch a page, pull out a date, write one sentence. A swarm of four agents is overkill. Invented cost count: each agent reads the same 8,000-token brief and adds 2,000 tokens of its own work. That is 4 times 10,000, or 40,000 tokens, before the agents send each other messages. One agent doing all three steps might use 10,000. Use several agents only when subtasks are independent and large enough to pay for the overhead.
Checking that reasoning helps
Reasoning models spend extra tokens thinking before they answer. Collect ten multi-step questions and ten one-step lookups. Run each set with reasoning on and off. Expect a clear gain on the multi-step set, little on the lookups, and a higher token bill both times. If you see no gain on your own task, switch it off.
Common mistake
Rejecting a whole idea because the marketing is silly. Look for the mechanism underneath.
A founder reads that "agentic AGI will replace developers". They could panic or ignore it. A better move: translate it. "Agentic" means a model that chooses its own steps. "AGI" has no agreed definition. "Replace developers" is a prediction, not a measurement. Translated, the sentence says almost nothing testable.
Check yourself
What question best tests a new buzzword?
Not quite. A useful term names something you can implement and measure.
When is a multi-agent system justified?
Not quite. Coordination cost makes it overkill for most tasks, but it fits parallel independent subtasks.
A study finds developers were 19 percent slower with AI tools. How should you treat this?
Not quite. The study had 16 developers and early-2025 tools, so weigh it as evidence, not a verdict.
You got 0 of 3 right.
Go deeper
- Article
METR · A randomized trial that found slower completion, with caveats.
- Article
Simon Willison · Draws the line between vibe coding and reviewed AI-assisted code.
- ArticleAI Snake OilArticle
Narayanan and Kapoor · A newsletter that separates real capability from claims.
Spotted a mistake or something unclear? Report it on GitHub.