The AI agent space has more jargon than the dot-com era did, and most of it is unnecessary. This glossary covers the 30 terms that show up in real product docs, sales calls, and procurement reviews — the ones you actually need to know to make a good decision. Every definition is one paragraph. No marketing fluff.
Use it as a reference. Save it. Send it to the person on your team who keeps asking what "MCP" means.
A
Agent
Software powered by a large language model that can take a goal in plain English, decide on the steps to reach it, and use tools (apps, the web, files) to finish the work. Differs from a chatbot in that it takes actions, not just messages.
Agentic AI
Marketing umbrella term for AI systems that act, not just respond. In practice, "agentic AI" and "AI agent" are used interchangeably. If a vendor uses "agentic" three times in a slide, watch closely — it often signals demo polish over working product.
API (Application Programming Interface)
The plumbing that lets one piece of software talk to another. Agents use APIs to connect to your CRM, calendar, email, and dozens of other systems. If your tool has no API, an agent can't talk to it.
Autonomous agent
An agent that can run end-to-end on a goal without a human approving every step. The opposite of a "human-in-the-loop" agent. Useful for low-risk tasks (research, drafting); risky for high-stakes ones (sending external emails, moving money).
C
Chain-of-thought
When a language model writes out its reasoning step by step before giving the final answer. Improves accuracy on complex tasks. Most modern agent frameworks use chain-of-thought under the hood whether you see it or not.
Context window
The amount of text a language model can read at once, measured in tokens. As of 2026, leading models handle 200,000 to 2 million tokens — roughly 150,000 words to a million-and-a-half. Bigger context = the agent can read your whole employee handbook in one shot.
Cowork
Anthropic's product that lets non-developers run Claude on their desktop with access to local files, MCP connectors, and skills. Used heavily in Build An Agent Day for knowledge workers automating their own jobs.
E
Eval (evaluation)
A test that measures how well an agent performs a specific task. Production agents have evals the same way good code has unit tests. If a vendor can't show you an eval suite for their agent, assume they don't have one.
Embedding
A way of turning text into a list of numbers that captures meaning. Used so an agent can search your documents by meaning, not just keyword. Powers most "talk to your docs" features.
F
Fine-tuning
Taking a general-purpose model and training it further on your specific data so it sounds more like you and knows more about your domain. In 2026, most teams skip fine-tuning and use skills plus RAG instead because both are cheaper and easier to update.
Function calling
The mechanism that lets a language model invoke a specific software function ("send_email," "create_invoice"). Function calling is what makes an agent able to do things rather than just describe them.
G
Guardrails
Rules that constrain what an agent can and cannot do. Examples: never send email to addresses outside our domain, always show me the draft before posting. Good guardrails are how you get autonomy without disasters.
H
Hallucination
When a language model produces information that sounds true but isn't. The fix is usually grounding (RAG) or tool use — make the agent fetch the real answer from a real source instead of guessing.
Human-in-the-loop (HITL)
A workflow design where a person approves key steps before the agent continues. Standard practice for any agent that touches money, customers, or compliance-sensitive work.
I
Inference
The act of running a language model to get an answer. Each inference call costs tokens, and tokens cost money. Designing efficient agents means minimizing wasteful inference.
L
LLM (large language model)
The text-prediction engine at the heart of every agent. Examples: Claude (Anthropic), GPT-5 (OpenAI), Gemini (Google), Llama (Meta). The LLM is the brain; tools and skills give it hands and a job description.
M
MCP (Model Context Protocol)
An open standard introduced by Anthropic in 2024 for connecting AI models to data sources and tools. MCP is to AI agents what USB was to consumer hardware — a common port that lets any agent talk to any tool. Most major agent platforms supported MCP by 2026.
Multimodal
A model that can read and produce more than just text — images, audio, video, PDFs. Most leading models are multimodal by 2026, which means an agent can read your screenshot, listen to your voicemail, or watch a short video and act on it.
O
Orchestration
Coordinating multiple agents or multiple steps to finish a complex job. Examples: a planning agent that spawns research agents, then hands their output to a writing agent. Used in serious enterprise deployments; usually overkill for a single user.
P
Prompt
The instructions you give a model. The art of writing good prompts is called "prompt engineering," though by 2026 the bar is lower because models follow plain English better than they did two years ago.
Prompt injection
A security attack where malicious instructions hidden in a webpage, email, or document trick the agent into doing something the user didn't intend. Real risk. Mitigations include input filtering, restricted tool access, and human approval for sensitive actions.
R
RAG (Retrieval-Augmented Generation)
A pattern where the agent first retrieves relevant information from a knowledge base, then generates an answer using that information. RAG is how most "AI that knows our company docs" actually works.
Reasoning model
A class of LLM trained to think longer and more carefully before answering. Slower and more expensive per call, but dramatically better on hard problems. Examples: Claude with extended thinking, OpenAI's o-series. Used selectively in agents — only when the task is hard enough to justify the cost.
S
Skill
A bundle of instructions, examples, and reference files that teaches an agent how to do a specific task the way your team does it. The 2025 breakthrough that made agents useful in real workplaces. Skills are written in plain English (with optional code) and stored as files, which means anyone can author one.
System prompt
The set of instructions an agent reads first, every time, that defines its role, rules, and personality. The system prompt is where you bake in your guardrails, your tone, and your approach.
T
Token
The smallest unit of text a model processes. Roughly four characters or three-quarters of a word. Pricing is per token, both for what you send (input) and what you receive (output).
Tool
A specific capability you give the agent — send an email, run a SQL query, create a calendar event, browse a webpage. The set of tools you grant an agent defines what it can actually do.
Tool call
A single use of a tool by the agent. Most modern agents make 5 to 50 tool calls to finish a real workflow.
V
Vector database
A specialized database that stores embeddings. Used in RAG systems so the agent can quickly find documents similar in meaning to the user's question. Examples: Pinecone, Weaviate, Chroma, pgvector.
W
Workflow
A repeatable business process — what your job actually is, broken down into steps. The whole point of building an agent is to take a workflow you already do and have the agent do it instead. The workshop runs on this idea.
How to use this glossary
Bookmark it. When a vendor or coworker drops a term you don't know, search this page first. If a term you need isn't here, email us at info@allthingsai.org and we'll add it.