AI Agents Glossary

How to use this glossary

Terms are in alphabetical order. If you are brand new, read these ten first: LLM, token, context window, prompt, hallucination, embedding, RAG, tool calling, AI agent and guardrails. Then follow the beginner's roadmap to learning AI.

Glossary: 72 AI agent terms

A2A (Agent2Agent protocol)

An open protocol, introduced by Google, that lets AI agents built by different teams or companies describe their abilities, discover each other and hand off tasks. Where MCP connects an agent to tools, A2A connects an agent to other agents.

Agent harness

The code around a model that turns it into an agent: the loop, the tool definitions, memory, safety checks and verification. Two agents using the same model can behave very differently because of their harness.

Agent loop

The repeating cycle at the heart of an agent: the model decides the next step, your code runs any tool it asks for, the result goes back to the model, and the cycle repeats until the task is done or a limit is reached.

Agentic AI

The approach of letting AI systems plan and take actions toward a goal rather than only generating a reply. See what is agentic AI.

AI agent

Software that uses a language model to decide and take actions toward a goal over several steps, using tools and memory. See what are AI agents.

API (application programming interface)

A defined way for one program to talk to another. You use a model provider's API to send prompts and receive responses from code.

API key

A secret string that identifies you to an API and is used for billing. Keep it in a .env file or secret manager — never in code you publish.

Chain

A fixed sequence of steps — for example prompt, then model, then output parser — run one after another. Chains follow a set path; agents choose their path.

Chunking

Splitting documents into smaller passages before embedding them for retrieval. Chunk size and overlap strongly affect how well RAG finds the right information.

Context engineering

Deciding exactly what goes into the model's context window on each call — instructions, retrieved documents, tool results, memory — so it has what it needs and nothing that distracts it.

Context window

The maximum amount of text, measured in tokens, a model can take into account in one call, including your prompt, history and its own reply. Anything outside it is invisible to the model.

Cosine similarity

A measure of how close two embeddings point in the same direction, commonly used to find text that is similar in meaning to a query.

Embedding

A list of numbers that represents the meaning of a piece of text, so that texts with similar meanings have similar numbers. Embeddings power semantic search and RAG.

Evaluation (evals)

Systematic testing of an AI system on a set of inputs with known good outcomes, run after every change. Evals are how you know an agent improved rather than got lucky once.

Few-shot prompting

Including a few worked examples of input and desired output in the prompt so the model copies the pattern.

Fine-tuning

Further training an existing model on your own examples to change its style or specialise it. Most agent projects need good prompts and retrieval long before they need fine-tuning.

Function calling

The original name for tool calling: the model returns a structured request to run a named function with arguments.

Grounding

Tying a model's answer to trusted sources — retrieved documents, database rows or tool results — rather than its memory, to reduce hallucinations.

Guardrails

Checks that keep an AI system safe and on-task: input filters, output validation, tool permissions, step and cost limits, and human approval for risky actions.

Hallucination

When a model states something false or invented with confidence. In agents a hallucination can become a wrong action, which is why grounding and verification matter.

Handoff

One agent passing a conversation or task to another, more specialised agent — for example from a general assistant to a billing agent.

Human-in-the-loop

A design where a person reviews or approves certain steps — such as payments or sending emails — before the agent continues.

Hybrid search

Retrieval that combines keyword search with semantic (embedding) search, which often finds better results than either alone.

Inference

Running a trained model to get an output. Every API call to a model is an inference request.

JSON schema

A formal description of the shape of JSON data — field names, types and allowed values. Used to define tool arguments and structured outputs.

Knowledge cutoff

The date after which a model has no training data. Anything newer must come from retrieval or tools.

LangChain

An open-source framework of building blocks for LLM apps — model interfaces, prompts, loaders, retrievers and integrations. See LangChain vs LangGraph.

LangGraph

An open-source framework for building agents as graphs of steps with shared state, supporting branching, retries, persistence and human approval.

Latency

How long a response takes. Agents that call several tools and models in sequence can be slow, so latency is a key design constraint.

LLM (large language model)

A neural network trained on huge amounts of text to predict the next token. It is the reasoning engine inside chatbots and AI agents.

Long-term memory

Information an agent keeps across separate conversations, usually stored in a database or vector store and retrieved when relevant.

MCP (Model Context Protocol)

An open standard, introduced by Anthropic, for connecting AI applications to tools and data. An MCP server exposes tools; any MCP-capable client or agent can use them. See the MCP guide.

MCP client

The part of an AI application that connects to MCP servers, lists their tools and calls them on the model's behalf.

MCP server

A program that exposes tools, data or prompts over the Model Context Protocol so AI applications can use them.

Model provider

A company or service that hosts models and sells access through an API, or the platform you run open models on.

Multi-agent system

Several agents, each with its own role and tools, working together on a task. Useful for complex work, but slower and costlier than a single agent.

Multimodal model

A model that works with more than text — images, audio, video or documents — as input or output.

Observability (tracing)

Recording every model call, tool call, input and output in a run so you can see why an agent behaved as it did and debug failures.

Open-weight model

A model whose trained weights are published so you can download and run it yourself, for example with Ollama.

Orchestration

Coordinating the steps, tools, state and agents in an AI system — the job of a loop or a framework such as LangGraph.

Parameters (model)

The learned numbers inside a neural network. Parameter count is a rough measure of model size, not a guarantee of quality.

Planning

An agent breaking a goal into smaller steps before or while acting on them.

Prompt

The input you give a model: instructions, context, examples and the question itself.

Prompt engineering

Designing prompts that reliably produce the output you need — clear roles, rules, examples and output formats.

Prompt injection

An attack where text inside a document, web page or message tries to give the model new instructions, such as leaking data or misusing a tool. Treat all tool output as untrusted data.

Quantization

Storing a model's numbers at lower precision so it runs faster and fits on smaller hardware, with a small loss in quality.

RAG (retrieval-augmented generation)

Fetching relevant passages from your own documents and giving them to the model so it answers from them. See RAG for AI agents.

ReAct

A prompting and agent pattern where the model alternates between reasoning about what to do and acting with a tool, then observes the result.

Reasoning model

A model trained to work through a problem step by step internally before answering, trading speed and cost for accuracy on harder tasks.

Reranking

A second retrieval step that re-scores candidate passages with a more precise model so the best ones reach the LLM.

Retrieval

Finding the most relevant information for a query from a document collection, database or search engine.

Router

A step that sends each request to the right model, tool or sub-agent based on what it asks.

Sandbox

An isolated environment where an agent can run code or commands without being able to harm the real system.

Semantic search

Search by meaning rather than exact words, usually using embeddings.

Short-term memory

The conversation history an agent keeps within a single task or session, sent back to the model on every call.

SLM (small language model)

A smaller, cheaper and faster language model, often good enough for narrow tasks and able to run on modest hardware.

State

The information an agent or workflow carries between steps — messages, intermediate results, flags. Frameworks like LangGraph make state explicit.

Streaming

Sending a model's output to the user token by token as it is generated, so answers appear faster.

Structured output

Model output constrained to a schema — usually JSON — so your code can use it directly instead of parsing free text.

System prompt

The instruction message that sets an agent's role, rules, tone and tool-use policy for the whole conversation.

Temperature

A setting that controls randomness in a model's output. Lower values give more consistent answers; higher values give more varied ones.

Token

The unit a model reads and writes — roughly a word piece. Prices, limits and context windows are all measured in tokens.

Tokenizer

The component that splits text into tokens and back. Different models use different tokenizers.

Tool

A function an agent may ask your code to run — search, a database query, a calculator, an API call — described to the model with a name, purpose and argument schema.

Tool calling

A model returning a structured request to use a tool with specific arguments, which your code executes and returns. It is what lets an agent act. See how to build an AI agent.

Top-p (nucleus sampling)

A setting that limits the model to the most likely tokens whose probabilities add up to p — another way to control randomness.

Transformer

The neural-network architecture behind modern LLMs, built around attention — a way for every token to weigh every other token in context.

Vector database

A database built to store embeddings and quickly find the ones most similar to a query. Used for RAG and semantic search.

Vibe coding

Building software by describing what you want to an AI coding assistant and iterating on its output. Productive for prototypes; still needs review and testing for anything real.

Voice agent

An agent you talk to: speech-to-text turns speech into text, the agent reasons and acts, and text-to-speech speaks the reply — for example a phone receptionist.

Workflow

A process whose steps are fixed in advance by the developer, even if some steps use an LLM. Contrast with an agent, which chooses its own steps.

Zero-shot prompting

Asking a model to do a task with instructions only and no examples.

Frequently asked questions

What is the difference between tool calling and MCP?

Tool calling is the model's ability to request a function. MCP is a standard way to package and serve those tools so any compatible agent can use them without custom integration code.

What is the difference between RAG and fine-tuning?

RAG gives the model relevant information at question time without changing the model; fine-tuning changes the model itself by training it further. For up-to-date or private knowledge, RAG is usually the right first choice.

What does 'agentic' mean?

Able to act toward a goal on its own — choosing steps, using tools and adapting to results — rather than only responding.