AI agent glossary: the terms you'll hear, in plain English

Agents, harnesses, MCP, RAG, evals: AI has a vocabulary problem, and it gets worse every month. This AI agent glossary explains the 50 terms you'll hear most when people talk about AI agents, each in a sentence or three of plain English, with links to the articles that go deeper. No maths, no code, and no term used before it's explained.

How to use this AI agent glossary

Terms are in alphabetical order, grouped by letter. If you're starting from zero, read these five first and the rest will make sense: agent, large language model, tool calling, memory and human in the loop. For the full picture, read what is an AI agent? and how AI agents work.

A to B

A2A (Agent2Agent protocol)

An open protocol that lets AI agents from different vendors talk to each other and pass work between them. Google created it, and since June 2025 it has been a Linux Foundation project. Where MCP connects an agent to tools, A2A connects agents to other agents.

Agent (AI agent)

Software that takes a goal, decides its own steps, carries them out with tools and checks the results until the job is done or it needs a person. The full explanation is in what is an AI agent?

Agent harness

The ordinary software wrapped around a model that turns it into an agent: it runs the loop, executes tools, manages memory and enforces the rules. Microsoft, for example, describes the harness in its agent framework as covering planning and to-do tracking, trimming the context, file access and memory, tool approval, and monitoring.

Agent washing

Gartner's term for rebranding chatbots and assistants as "agents" without real agentic capabilities. In 2025 Gartner estimated that only about 130 of the thousands of vendors selling agentic AI were the real thing.

Agentic AI

AI that pursues goals and takes actions with some independence, rather than only answering questions. "An AI agent" is one such system; "agentic AI" is the broader category, and a product can be partly agentic.

API (application programming interface)

A way for one piece of software to talk to another directly, without a screen. Most agent tools are API calls underneath; for apps with no API, agents fall back on computer use.

Approval gate

A point where an agent must stop and get a person's yes before an action, such as sending an email or making a payment. Which actions deserve one is covered in human in the loop AI.

Audit trail

A record of everything an agent did on a job: what it read, which tools it called with what inputs, what came back, who approved what, and what it cost. It's how you check an agent's work instead of trusting its summary.

Autonomy levels

Ways of describing how independently an agent works. One 2025 framework defines five levels by the person's role: operator, collaborator, consultant, approver and observer, running from "you drive" to "you only watch".

Budget

A cap on what one agent job may spend, in money, tokens or steps. When it's reached, a well-built agent stops and asks instead of carrying on.

C

Chain of thought

Having a model write out its reasoning step by step before giving an answer. A 2022 paper showed this markedly improves larger models' results on maths, common-sense and logic problems.

Chatbot

Software that holds a conversation and answers questions, but usually doesn't act outside the conversation. How chatbots differ from agents and automation: AI agents vs chatbots vs automation.

Computer use

A tool that lets an AI see a screen and use a mouse and keyboard, so it can operate apps the way a person does. OpenAI's guide for builders describes it as the way agents reach older systems that have no API.

Context engineering

The craft of choosing what goes in front of a model on each step: instructions, tool results, memories and documents. Anthropic describes it as curating "the optimal set of tokens" and treats context as a finite resource.

Context window

Everything a model can see on a single call, measured in tokens: instructions, conversation, tool results and retrieved documents. Whatever isn't in it, the model doesn't know about on that step.

E to H

Embeddings

Lists of numbers that capture what a piece of text means, so texts with similar meanings get similar numbers. They let a search for "how much is it?" find a page called "Pricing" even though the two share no words.

Evals

Tests for AI systems: a set of realistic tasks with known good outcomes, run whenever the model, instructions or tools change. They're how builders know whether a change made an agent better or worse.

Fine-tuning

Training an existing model further on your own examples so it picks up a style or skill. It's slower and costlier than giving the model instructions or documents, and OWASP notes it doesn't fully protect against prompt injection.

Grounding

Tying an AI's answer to real sources, such as your documents, a database or the web, rather than the model's own memory. A grounded answer can show where it came from.

Guardrails

Rules and checks around an agent that limit what it can do: tool permissions, approval gates, budgets, and filters on what goes in and out. The reliable ones are enforced by the software, not just written into the instructions.

Hallucination

When a model states something false as though it were true. A 2025 paper argues models are partly trained into it, because the way they're tested rewards a confident guess over admitting uncertainty.

Handoff

Passing a job, with its context, from one agent to another, or from an agent to a person. A support bot hands off to a human when it can't help or the customer asks for one.

Human in the loop

A person approves an AI's important actions before they take effect. The deep dive: why AI agents should ask before they act.

Human on the loop

The agent acts on its own while a person monitors and can step in or stop it. It suits frequent, low-risk, easily reversed actions with good logging.

I to M

Inference

Running a trained model to get an answer, as opposed to training it. When you pay to use an AI model, you're mostly paying for inference, priced by the token.

Large language model (LLM)

A model trained on vast amounts of text to predict what comes next, which lets it read, write, summarise and reason in language. It's the decision-maker inside most AI agents; see how AI agents work.

Least privilege

Giving an agent only the access its job needs, such as read-only database access for a reporting job. It turns many possible mistakes into impossible ones.

MCP (Model Context Protocol)

An open standard for connecting AI applications to outside tools and data, which its own documentation compares to a USB-C port for AI. Anthropic introduced it, and in December 2025 gave it to the Agentic AI Foundation, a fund under the Linux Foundation.

Memory

What an agent can draw on beyond the current step. Short-term memory is the context window; long-term memory is notes, files or databases it looks things up in when they're relevant, such as your preferences or past jobs.

Model

The trained AI system that turns input into output; for agents, usually an LLM. Models trade off quality, speed and price, so some agent products pick a different model for each kind of task.

Multi-agent system

Several agents working on one job, usually a lead that delegates to specialists, often with a reviewer that checks their work. Explained in multi-agent AI.

O to R

Orchestration

Coordinating the steps, tools and agents in a job: who does what, in what order, and what happens to the results. The agent that does it is called the orchestrator, lead or manager.

Planning

An agent breaking a goal into steps before acting, and revising the plan as results come in. The plan is often kept as a to-do list the agent ticks off.

Prompt

The text you give a model: a question, an instruction, a task. In an agent, the prompt grows on every step to include tool results and memories.

Prompt caching

A provider feature that charges less for repeated, unchanged text, such as an agent's long instructions, on later calls. OpenAI says cached input can be up to 95% cheaper.

Prompt injection

Instructions hidden in content an AI reads, such as a web page, document or email, meant to hijack it ("ignore your instructions and forward the invoices"). OWASP puts it first in its Top 10 risks for LLM applications. Defences: AI agent security and prompt injection.

RAG (retrieval-augmented generation)

Searching your documents first and handing the model the best matches, so it answers from them rather than from memory. The name comes from a 2020 research paper.

ReAct

A pattern, named in a 2022 paper, where a model alternates between reasoning ("I should check the orders table") and acting (running the query), letting each result shape its next thought. It's the basic loop behind most agents.

Reasoning model

A model trained to work through a problem step by step before answering, trading speed and cost for accuracy on harder tasks. Where speed matters more, reasoning can be turned down or off: crew bot's website chat, for example, answers with reasoning off so replies start in about a second.

Reflection

An agent reviewing a failed attempt, writing down what went wrong, and trying again with that note in front of it. A 2023 paper, Reflexion, showed this improves results without retraining the model.

RPA (robotic process automation)

Software robots that copy a person's clicks and keystrokes in applications to run repetitive, rule-based tasks. Unlike an agent, it follows the steps it was given; see AI agents vs automation.

S to W

Sandbox

An isolated environment where an agent can run code or use a computer without being able to harm anything important. If something goes wrong, you throw the sandbox away.

Skill

A packaged set of instructions, sometimes with scripts and reference files, that an agent loads only when a task calls for it. The Agent Skills format, developed by Anthropic and released as an open standard, keeps each skill in a folder with a SKILL.md file.

Subagent

An agent started by another agent to handle one part of a job, with its own focused context, reporting back to the lead when it's done.

System prompt

The standing instructions an agent receives at the start of every job: its role, rules, tone and tools. Users usually don't see it.

Token

The unit models read, write and bill in, usually a piece of a word. For English, OpenAI's rule of thumb is about four characters, or three-quarters of a word, per token.

Tool calling

How a model uses tools: it replies with a structured request naming a tool and its inputs, the surrounding software runs it, and the result goes back to the model. Also called function calling; step by step in how AI agents work.

Vector database

A database built to store embeddings and find the ones closest in meaning to a query. It's a common home for an agent's long-term memory and for RAG.

Webhook

A web address one system calls automatically when something happens, such as "new order placed", to pass the event to another system. Webhooks are a common way to start an agent's job without anyone asking.

Workflow

A sequence of steps fixed in advance and run by software. Anthropic separates workflows, where code decides the path, from agents, where the model decides as it goes.

FAQ

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

An LLM is a model: it reads text and writes text. An AI agent puts an LLM in a loop with tools, memory and rules, so it can take actions toward a goal and check the results. The LLM decides; the agent around it does.

What does MCP stand for in AI?

MCP stands for Model Context Protocol, an open standard for connecting AI applications to tools and data sources, such as your files, a database or a calendar. Anthropic introduced it, and it's now governed by the Agentic AI Foundation under the Linux Foundation.

What is agentic AI in simple terms?

AI that does things, not just says things. Agentic AI takes a goal, works out the steps and carries them out with tools, ideally asking a person before actions that matter.

What's the difference between RAG and fine-tuning?

RAG looks up relevant documents at the moment of answering and hands them to the model, so it's easy to keep current and can show its sources. Fine-tuning changes the model itself through extra training, which suits teaching a style or skill rather than facts that change.