The AI agents glossary
Every core term in agentic AI, defined in plain English — with how it works, a concrete example, and a link to the deep dive. Bookmark it as your reference for building and writing about AI agents.
- 20+ terms defined
- Plain English
- Updated 2026
Summary
In summary, This glossary defines the core terms of agentic AI in plain English, from AI agent and large language model to RAG, tool calling, orchestration, and guardrails. Each entry gives a precise definition with a concrete example and links to a deep-dive guide. Terms are grouped into foundations, reasoning and tools, knowledge and memory, and systems and safety.
New fields invent new vocabulary, and agentic AI has invented a lot of it fast. This glossary cuts through the jargon: each entry gives a precise, jargon-light definition of a term you'll meet while building or evaluating AI agents, then links to a full guide when you want depth. Start anywhere — the terms below are grouped so related concepts sit together.
Core concepts
The building blocks — what an agent is, the model that powers it, and the limits it works within.
- AI agent
- Software that perceives, reasons, plans, calls tools, and acts autonomously toward a goal in a loop.
- Agentic AI
- The paradigm of AI systems that act with agency — pursuing goals via reasoning and tools, not just generating output.
- Large language model (LLM)
- A neural network trained on vast text that predicts tokens — the reasoning engine inside most agents.
- Inference
- Running a trained model to generate outputs from inputs; the cost and latency of every model call.
- Context window
- The maximum number of tokens a model can consider at once — a key constraint and cost driver for agents.
- Autonomous agent
- An agent that pursues goals with minimal human steering — see the levels-of-autonomy guide.
How agents think and act
The patterns and mechanisms that turn a language model into something that reasons and takes action.
- ReAct (Reasoning + Acting)
- A pattern that interleaves Thought → Action → Observation so an agent reasons, acts, and re-reasons.
- Chain-of-thought
- Prompting a model to reason step by step before answering, improving multi-step accuracy.
- Prompt engineering
- Designing prompts and system instructions to steer model behavior reliably.
- Hallucination
- When a model produces plausible but false or unsupported information; reduced with grounding and verification.
- Tool calling
- An agent's ability to invoke external tools and APIs to take actions and fetch data.
- Function calling
- The mechanism where an LLM emits a structured call (name + JSON args) the runtime executes.
- JSON Schema
- The standard vocabulary for declaring a tool's parameters — types, required fields, enums — so the model fills them and the runtime validates them.
How agents know things
Grounding answers in real data and remembering across steps and sessions.
- Retrieval-augmented generation (RAG)
- Retrieving relevant documents and adding them to the prompt so the model answers from real, current data.
- Embeddings
- Numeric vector representations of text that place similar meaning close together, enabling semantic search.
- Vector database
- A store of embeddings that retrieves the most similar ones via nearest-neighbor search.
- Agent memory
- How an agent retains information — short-term context and long-term stores — to stay coherent over time.
- Fine-tuning
- Further training a pretrained model on task or domain data to adjust its behavior, style, or skills.
Coordinating and controlling agents
Putting agents together into systems — and keeping them safe and on-policy.
- Multi-agent system
- Multiple specialized agents that coordinate — delegating, reviewing, and handing off — to solve a goal.
- Orchestration
- Coordinating control flow between an agent's steps or between agents — routing, sequencing, retries.
- Guardrails
- Safety controls that constrain agent inputs, outputs, and actions to keep behavior safe and on-policy.
- Model Context Protocol (MCP)
- An open standard for connecting models and agents to external tools and data through a uniform interface.
Want the full picture?
Each definition links to a deep-dive guide. If you're starting from scratch, read what is agentic AI first, then follow the links from there. Comparing tools or approaches? The comparisons hub puts the trade-offs side by side.
About this glossary
It's a free, plain-English reference that defines the core terms behind agentic AI — from 'AI agent' and 'agentic AI' to RAG, embeddings, tool calling, orchestration, and guardrails. Each entry gives a precise definition, explains how the concept works, shows a concrete example, and links to the deep-dive guide so you can go from a quick definition to a full understanding.
From terms to a working agent
You know the vocabulary — now build the thing. Start free, no credit card required, and ship your first agent in minutes.