AI / Agent Terminology Index

This is an index of terms with the vendor-specific name for the same concept noted where it applies. Built with the help of Claude.

---

1. Core model concepts

LLM (Large Language Model): the underlying model (Claude, GPT, Gemini) that generates text by predicting likely next tokens.

Foundation model: a large, general-purpose model trained on broad data, meant to be adapted to many tasks rather than one.

Token: the unit a model reads/writes text in; roughly ¾ of a word in English. Pricing, context limits, and speed are usually measured in tokens, not words or characters.

Context window: how much text (conversation history, documents, tool results) the model can “see” at once during a single conversation. Resets when the conversation ends.

System prompt / instructions: standing instructions given to the model before the conversation starts, defining its role, tone, or constraints. Every vendor has this

Prompt: the input you give the model. “Prompt engineering” is the practice of wording that input for a better result.

Few-shot prompting: giving the model a few examples of the desired input/output pattern inside the prompt itself, rather than relying on its general training.

Fine-tuning: retraining a copy of a model on your own data so its default behavior shifts. Distinct from memory or context; fine-tuning changes the model itself, permanently, and is a developer-level operation, not something available in a chat window.

Hallucination: the model generating a confident but false or fabricated statement.

Embeddings: a numeric representation of text used for similarity search (e.g., “find posts related to this topic”). Underlies most recommendation and search features, including RAG below.

RAG (Retrieval-Augmented Generation): the model retrieves relevant documents/data before answering, rather than answering purely from its training. Web search and connector lookups are both forms of RAG in practice.

---

2. Connecting AI to outside systems

API (Application Programming Interface): the general term for a system exposing functions other software can call. Predates AI; every vendor’s AI integrations sit on top of ordinary APIs.

Webhook: a way for one system to notify another automatically when something happens (“new email arrived”), rather than being polled. Common in no-code automation (Zapier/Make) and increasingly in agent triggers.

OAuth: the standard authentication handshake (“Sign in with Google”) used when you connect a service to an AI tool. What “connect” actually does under the hood.

The same concept, different vendor names

Concept	Anthropic (Claude)	OpenAI (ChatGPT)	Google (Gemini)
Standard for AI ↔ external system	MCP (Model Context Protocol) — open standard, adopted by OpenAI, Google, AWS, Microsoft	Function calling (the underlying mechanism); MCP also supported	Function calling; MCP also supported
A specific integration to one service ie: WordPress	Connector	GPT Action (inside a Custom GPT)	Extension
A callable capability inside that integration ie: Write a Post	Tool	Function	Function
A saved, specialized configuration/persona	(Project custom instructions)	Custom GPT	Gem

Scope / permissions: what a specific tool/function/action is allowed to do, set when you authenticate it. read vs. write permissions

---

3. Agents and automation

Agent / agentic: not a distinct product, but a mode: the model makes a sequence of tool calls, reads each result, and decides the next step itself, rather than answering once.

Orchestration / multi-agent: a setup where multiple agents (or agent + sub-agents) coordinate on parts of a task, rather than one model doing everything.

Tool call: one specific instance of an agent invoking a tool/function/action mid-task.

Autonomy level: how much an agent does before checking in with a person; a design choice (see guardrails below), not a fixed technical property.

RPA (Robotic Process Automation): older, pre-AI automation that follows fixed, deterministic rules (click here, type this); an AI agent is worse at perfect repetition but handles judgment calls RPA can’t.

No-code automation layer: Zapier, Make, n8n. Their unit of work is called a “Zap” (Zapier) or “Scenario” (Make): a trigger plus a chain of actions, increasingly with an AI/agent step inserted into that chain rather than being the whole workflow.

---

4. Control and safety terms

Guardrail: any constraint placed on an agent’s behavior: requiring confirmation before an action, restricting which tools/functions it can use, or hard-coded rules it must follow. Not a built-in feature – something you design.

Human-in-the-loop: a guardrail specifically requiring a person’s approval before an action executes. Where you require it (every action vs. only high-risk writes) is the actual design decision.

Memory: information that persists across separate conversations (facts saved about you or your business, recalled later). Distinct from the context window, which only holds the current conversation.

Session: one continuous conversation/interaction; ends when the context window resets.

---