A Plain-English Glossary of AI Assistant Terms
Maker's note: 기준일 2026-10-03. We're the AZET team — we build and operate azet.io — and this glossary covers the terms that appear across this blog and in AI product discussions generally. Definitions are plain-English first, precise second. AZET-specific notes are marked and dated.
Every field accretes jargon faster than it explains it, and AI assistants have compressed a decade's worth of new vocabulary into about three years. The result is confident nodding at words half-understood. This glossary is the one we hand people when they ask "what does that actually mean?"
The terms
| Term | What it actually means |
|---|---|
| AI assistant | A product built around a language model that you instruct in plain language — it answers, drafts, summarizes, and sometimes carries out multi-step tasks. The product, not just the model. |
| Agent | An AI system that takes a goal and decides its own steps toward it — opening pages, running tools, checking results — rather than answering in one shot. More autonomy, more need for verification. |
| Agentic browsing | An agent navigating the live web for you: reading pages, clicking through, extracting what it finds. Distinct from answering from memory, and only as trustworthy as its reading. |
| Browser automation | A script driving a browser through steps a person wrote in advance. Rigid by design — the same chore runs the same way every time, and success is checkable. |
| Headless browser | A browser running without a visible window, so automation can work on pages without rendering anything to a screen. Same engine, no face. |
| Model | The trained system that actually generates text — the engine. "Assistant" is the car around it: instructions, tools, memory, safety rails. |
| Prompt | Whatever you type to the assistant. Writing a clearer prompt is the highest-leverage skill in using one; most "the AI is bad at this" is really "the prompt didn't say what was wanted." |
| Token | The small chunks of text (roughly a word or a fragment) that models read and write in. Usage, limits, and API pricing are counted in tokens, not words. |
| Context window | Everything the model can currently "see" — your prompts, its answers, pasted text, and pages it has read. Exceed it and earlier content falls out of view. Not long-term memory. |
| Session | One working conversation or run, with its own accumulated context. In agent tools, sessions can be paused and resumed with plan and progress kept. |
| Checkpoint | A saved state of a run — plan, position, pending instructions — that can be returned to. The difference between "undo" and "start over." |
| Hallucination | A fluent, confident output that is simply not grounded in anything true. Not lying (no intent), not a crash — it reads exactly like a correct answer. |
| Verification receipt | Evidence tied to a claim: version, hashes, observation date, scope. In our project, even a passed test is read with its receipt — never the verdict alone. |
| Deterministic | Same input, same steps, same output, every time. Browser automation is deterministic; a language model choosing its own route is not. |
| Hosted model | The model runs on a provider's servers; your text travels to it. Strongest capability, zero setup, and privacy depends on the provider's policies. |
| Local model | The model runs on your own hardware; text stays home. You pay in setup, hardware, and generally lighter capability. |
| MCP | Model Context Protocol — an open standard for connecting models to external tools and data sources, so an assistant can use a tool without custom integration work per tool. |
| Waitlist | A queue for early access: register an email, get in line, hear when it's your turn. Signals a product under construction rather than generally available. |
| Credits | Prepaid units of usage inside a product's own accounting — on azet.io, 1 credit equals $0.01, per our project README as of October 2026. |
| Notarization | Apple's process of checking and signing macOS software so Gatekeeper opens it without frightening warnings. Worth knowing as a buyer: unsigned or ad-hoc software hasn't passed it. |
Three distinctions worth memorizing
If the table is too much, keep these three separations — they resolve most confusion:
Assistant vs agent. An assistant answers and drafts within a conversation; an agent takes a goal and runs steps autonomously. The distinction predicts failure modes: assistants stall, agents wander.
Memory vs context. A model that recalls your last five pages of conversation is not remembering you — that's the context window, and it empties. "Memory" means the product explicitly stores and re-injects facts, and it should say so.
Automation vs agency. A script does exactly what was written, identically every time. An agent improvises a route to the goal. Automation is auditable; agency is flexible. Knowing which one you're holding is most of supervising it.
Why we keep a glossary at all
Because vocabulary is where products quietly overclaim. "It remembers you" (does it?), "it verifies answers" (against what?), "it browses autonomously" (who approved the clicks?). Precise terms make vague marketing visible.
FAQ
What is the difference between an AI assistant and an AI agent? An assistant responds within a conversation — answers, drafts, summaries. An agent takes a goal and decides its own steps toward it, including using tools and browsing. Assistants stall; agents wander.
What is a token in simple terms? The small chunks of text a model reads and writes in — roughly a word or a piece of one. Usage limits and API pricing are measured in tokens, which is why long pastes cost more than short questions.
What does "context window" mean, and is it memory? It's everything the model can currently see: your prompts, its replies, pasted text, and pages read during the session. It is not memory — exceed it and earlier content drops out of view.
What is MCP? Model Context Protocol: an open standard for connecting models to external tools and data, so an assistant can use tools without bespoke integration per tool. Think of it as a common plug shape for capabilities.
Do these terms mean the same thing on azet.io? The category terms do. AZET-specific meanings — like 1 credit equaling $0.01 — are marked in the table and dated October 2026, per our project README; azet.io carries anything newer.