AI Agent Glossary: Key Terms Explained Simply for Everyone

This AI agent glossary explains the terms you will meet when reading about or working with AI agents, in plain language and without jargon piled on jargon. Terms are grouped by theme, so you can read straight through or jump to the section you need.

If you are brand new to the topic, start with our guide to what an AI agent is, then come back here whenever a word trips you up.

The basics

AI agent: Software that uses an AI model to pursue a goal on its own. It decides the next step, uses tools to act, checks the results, and repeats until the task is finished.

Large language model (LLM): The AI model at the core of most agents. It is trained on huge amounts of text and can read, write, reason about, and generate language and code. Claude, GPT, and Gemini are families of LLMs.

Chatbot: A conversational interface that responds to messages one at a time. Unlike an agent, it usually does not take actions or work through multiple steps on its own.

Prompt: The instruction or input you give an AI model. For agents, the prompt often includes the goal, background context, and rules to follow.

System prompt: Hidden or behind-the-scenes instructions that set an AI’s role, tone, and rules before the conversation starts. Agents usually run with a detailed system prompt describing their tools and limits.

Token: The small chunk of text a model reads and writes, often a word or part of a word. Model usage is commonly measured and billed in tokens.

Context window: The maximum amount of text, measured in tokens, a model can consider at once. It includes instructions, conversation history, file contents, and tool results. When a task outgrows it, older details may be summarized or dropped.

Inference: The act of running a trained model to produce an output. Every step an agent takes involves at least one round of inference.

Hallucination: When a model states something false with confidence, such as an invented function name or a made-up fact. Tool use, tests, and review help catch hallucinations.

How agents work

Agent loop: The repeating cycle at the heart of every agent: plan, act, observe, and adjust. The loop continues until the goal is met or the agent stops for help.

Agentic workflow: A process where AI works through a task in steps instead of giving one answer. Read more in agentic workflows explained.

Tool use (function calling): The ability of a model to call external functions, such as reading a file, running a command, or searching the web. Tools are what let an agent act rather than only talk.

Planning: The step where an agent breaks a goal into smaller tasks before acting. A written plan reduces wandering and makes work easier to review.

Reflection: When an agent evaluates its own output and decides whether to fix something. Running tests and rereading results are common forms of reflection.

Reasoning: The model’s step-by-step thinking about a problem before answering or acting. Some models can be set to spend more effort reasoning on hard tasks.

Effort: A setting some tools offer to control how much thinking a model does per step. Higher effort can improve quality on complex tasks but takes longer and costs more.

Memory: Information an agent keeps between steps or sessions, such as notes, past decisions, or project facts. Short-term memory lives in the context window; long-term memory is usually stored in files or a database.

Autonomy: How much an agent decides on its own versus asking a human. More autonomy means less supervision, which makes guardrails more important.

Tools, skills, and connections

MCP (Model Context Protocol): An open standard introduced by Anthropic for connecting AI apps to tools and data sources. It gives agents a consistent way to reach things like files, databases, and services. Learn more in what is MCP.

MCP server: A program that exposes a set of tools or data through MCP, so any compatible agent can use them.

Agent Skills: Folders containing a SKILL.md file, with a name and description at the top followed by instructions and optionally scripts or resources. Agents load a skill when it is relevant to the task. The format was introduced by Anthropic and has been adopted more widely.

API: An interface that lets software talk to other software. Agents often call APIs to fetch data or take actions in other services, and many AI models are themselves accessed through an API.

API key: A secret string that identifies you to a service and authorizes usage. API keys should never be pasted into public places or committed to code repositories.

RAG (retrieval-augmented generation): A technique where relevant documents are looked up and added to the model’s context before it answers. It helps the model use specific, current information it was not trained on.

Coding agents

Coding agent: An agent specialized in software work: reading a codebase, editing files, running commands and tests. Examples include Claude Code, Codex, Gemini CLI, and OpenCode.

CLI (command-line interface): A text-based way to run programs in a terminal. Many coding agents run as CLI tools inside your project folder.

Repository (repo): A project’s folder of code tracked by version control, usually Git. Agents typically work inside a repo so every change can be reviewed and undone.

Tests and checks: Automated programs that verify code works as expected, such as unit tests, linters, and type checkers. They give agents an objective signal of success.

Merge: Combining changes from one line of work into the main project. In multi-agent setups, reviewed work is merged once it passes checks.

Vibe coding: A casual style of building software by describing what you want to an AI and accepting its output with little manual coding.

Teams of agents

Multi-agent system: A setup where several agents with different roles work together, such as planner, builder, and reviewer. It helps with large tasks and catches more mistakes.

Team lead agent (orchestrator): An agent that breaks a goal into tasks, assigns them to other agents, and coordinates the result, much like a project manager.

Handoff: Passing a task, along with its context, from one agent to another. Clear handoffs keep a team from duplicating or losing work.

Code review agent: An agent that reads another agent’s changes and checks them for bugs, style, and fit with the project before they are accepted.

Safety and control

Human in the loop: A design where people approve key steps, such as a plan or final result, rather than letting the agent run fully unsupervised.

Permissions: Rules defining what an agent may and may not do, such as which folders it can edit or which commands it can run. Good tools keep permissions as narrow as the task allows.

Guardrails: Safety limits that stop an agent from harmful actions, such as blocking risky commands or protecting sensitive files.

Sandbox: An isolated environment where an agent can run code without affecting the rest of your system.

Models and where they run

Cloud model: A model run by a provider on its own servers and reached over the internet, usually with an account or API key.

Local model: A model that runs on your own computer, so your data does not leave the machine. Tools like Ollama make this easier. See running AI models locally with Ollama.

Open-weight model: A model whose trained parameters are published so anyone can download and run it, often locally.

FAQ

What are the most important AI agent terms to know?

Start with AI agent, LLM, prompt, context window, tool use, and agent loop. Those six cover how agents think and act. Once they make sense, terms like MCP, agent skills, and multi-agent systems are much easier to follow.

What is the difference between an AI model and an AI agent?

The model is the brain that reads and generates text. The agent is the system around it that gives the model a goal, tools, and a loop to work through steps. The same model can power a chatbot or an agent.

What does “agentic” mean in AI?

“Agentic” describes AI that acts with some independence: it plans, takes actions, and adjusts based on results instead of just answering one question. An agentic workflow is a task carried out this way.

What is the difference between MCP and agent skills?

MCP connects an agent to external tools and data sources through a standard protocol. Agent skills are packaged instructions, and sometimes scripts, that teach an agent how to do a particular kind of task. They work well together.