What we mean
These are the words you’ll hear in week one of going AI-native. We keep them in one place so the conversation can go faster — not a glossary to memorize, just a reference so nobody has to stop and ask what something means in the middle of the work.
Words are infrastructure too. A team that means the same thing by agent moves faster than a team that doesn’t, for the same reason a team with shared skills moves faster than a team with passed-around prompts.
01 — Basics
What it thinks in
The raw material under everything else.
Token
The unit of text the model breaks language into, roughly three-quarters of a word. Models read, generate, and price in tokens. “Use less context” really means “spend fewer tokens.”
Context
Everything the model can see when it answers: your message, the system prompt, files, tool outputs, the running history. Its working memory for this turn.
Context window
How much context the model can hold at once. When it fills, older content gets dropped or summarized. Cleaner context beats more context.
Prompt
What you tell the model. Most “the AI got it wrong” moments are actually prompt problems.
02 — Building blocks
What extends it
The pieces that turn a chat model into something that can do work.
Skill
A folder with a SKILL.md file that teaches the AI how to do a specific kind of task. Reusable, loaded on demand. (This post was written by one.)
Tool
A function the AI can call: read a file, run code, send an email. Tools are how an AI does anything beyond generating text.
MCP
The standard for connecting outside services to an AI. An MCP server exposes tools for the AI to call.
Connector
A specific MCP wired up to a service. “I connected Slack” means you added the Slack connector.
Subagent
A child agent spawned to do a focused subtask in its own context. Used to keep the main agent’s context clean and to parallelize.
03 — Actors
What does the work
The thing that’s actually moving.
Agent
An AI that pursues a goal across multiple steps: reads the situation, picks tools, acts, checks its work, adjusts. A chatbot answers your question. An agent goes and does the thing.
Assistant
Loose, friendly term for AI that helps a person. When precision matters, prefer agent or just the model.
Copilot
An in-app AI that suggests as you work, or specifically GitHub Copilot. Its DNA is autocomplete, not autonomy.
04 — Surfaces
Where it runs
The product names — the chat apps, the terminal coding agents, the desktop agents. This is the part of the vocabulary that churns fastest, so we keep it short: when someone names a product, ask which layer it lives at (chat, coding agent, desktop agent) and you’ll usually have what you need.
05 — Outputs
What you keep
The thing that survives the session.
Artifact
A persisted output you can come back to: a dashboard, a tracker, a report that survives when the chat ends.
Document / file
Plain output to disk. Less interactive, more portable. Often the right answer when the deliverable is “send me the thing.”
06 — Practices
How to use it well
The how, not the what.
Agentic
Describes AI that takes initiative, decides, uses tools, and works toward a goal without step-by-step prompting. The opposite of a single-shot chat.
Enablement
The work of getting your team actually using AI well: training, examples, shared skills, removed blockers. The difference between paying for licenses and people using them by Friday.
Context engineering
Getting the right information into the context window at the right time, and keeping junk out. Often matters more than prompt engineering on serious work.
RAG
Pulling relevant docs from your own data into the model’s context so it can ground its answer in actual facts.
07 — Watchouts
How to keep it honest
The failure modes that show up the moment an agent has real authority.
Guardrails
Explicit limits on what an agent can do. Guardrails are how you trust an agent enough to give it more rope.
Approval gate
The agent prepares an action but waits for a human to approve before executing. The most common pattern for getting comfortable with agents in production.
Hallucination
Plausible-sounding output that’s wrong. The model isn’t lying; it doesn’t know it doesn’t know.
Drift
When a system’s behavior gradually moves away from what you set up. The wiki goes stale, the skill stops matching reality. Drift is why these systems need maintenance, not just install.
If you want your agents to share this vocabulary — so the words mean the same thing whether a person or a model is reading — hand them this page:
Give your workspace this vocabulary
- An AI workspace with persistent context (CLAUDE.md, AGENTS.md, or memory)
Read https://obvios.co/practice/what-we-mean/ (or this post's text if it
is already in your context). Extract every term and its definition,
keeping the definitions intact — do not paraphrase them more loosely.
Add a "Shared vocabulary" section to this workspace's persistent context
(CLAUDE.md, AGENTS.md, or memory, whichever this workspace uses). Use
these meanings when you read and write for this team. If a term already
has a different local meaning here, keep the local meaning and note the
difference next to it — do not overwrite it.
Cite the source URL at the top of the section so future agents know
where the vocabulary came from and where updates live.
The list will change. Drift applies to glossaries too — when our words change, this page changes with them.