An AI agent is software that carries out tasks for you, rather than just answering your questions. The chat window you already use is the answering kind: you ask, it replies. An agent is the doing kind: it books the appointment, sends the email, updates the records. The twelve terms below make up the working vocabulary that has grown up around these agents in 2026. You have probably heard most of them. What follows is what each one actually means, and where it shows up in practice.
A few months ago we published a glossary of 50+ AI terms worth knowing. This is the focused sequel, narrower on purpose. The center of gravity in AI has shifted from what a chatbot can tell you to what an agent can go and do, and that shift brought its own set of words with it.
What is MCP?
MCP, short for Model Context Protocol, is a shared standard for connecting AI agents to the tools and data they work with. The name is a mouthful; the idea is simple. Think of how every gadget used to ship with its own charger, until USB-C arrived and one plug fit nearly everything. MCP is that, but for software: instead of a developer hand-building a custom connector between each AI and each app, there is one standard they all speak. When you hear that a tool “supports MCP,” it means an agent can plug into it without that custom wiring. The practical result is that connecting AI to the software you already run is getting dramatically faster and cheaper. We went deeper on this in MCP versus a plain API.
What is the agent loop?
The agent loop is the repeating cycle an agent runs to get something done: perceive, plan, act, observe. It takes in the current situation, decides on a next step, takes that step, checks the result, and goes around again. This is the mechanical difference between a chatbot and an agent. A chatbot answers once and stops. An agent keeps cycling until the task is finished or it runs out of road. Nearly everything else on this list is either a way to make that loop smarter or a way to keep it from causing harm.
What is tool use?
Tool use is an agent’s ability to take real actions, not merely describe them. On its own, a language model can only produce text. It can write you a flawless email and then do nothing further, because writing is all it can do. Give it tools, meaning access to an inbox, a browser, a calendar, a database, and the gap closes: now it can actually send the email, look the fact up, book the meeting. Tool use is the dividing line between AI that advises you and AI that does the thing for you.
What is an orchestrator?
An orchestrator is the agent that manages other agents. Some goals are too large or too varied to be a single task. “Launch the spring promotion” is not one job; it is writing the emails, scheduling the posts, updating the website, and building the discount codes, all at once. An orchestrator takes a goal of that shape, breaks it into pieces, hands each piece to whichever agent is suited to it, and assembles the results. It is less a worker than a manager. Its real job is deciding what needs doing and who should do it.
What is a subagent?
A subagent is a specialist: one agent assigned a single, narrow job. When an orchestrator divides a large goal, each slice goes to a subagent built to handle just that slice. The logic is the same reason a firm hires specialists instead of asking one person to be its lawyer, its accountant, and its designer all at once. An agent with a narrow task and little to distract it produces better, more reliable work than a generalist trying to juggle everything. Narrow and deep tends to beat wide and shallow, in software as in offices.
What is agent memory?
Agent memory is how an agent holds on to information beyond the current conversation. It comes in two layers. Short-term memory is what the agent has in front of it right now, the contents of the exchange you are having. Long-term memory is the part that persists: a durable store the agent can write to and read back later, so that facts survive after the chat window closes.
The second layer is the one that matters most in practice, and it is easiest to appreciate by its absence. Anyone who has leaned on AI for real work has hit the moment where you are re-explaining the same context for the tenth time, which systems you use, how your operation is set up, the preferences you have stated again and again. Without long-term memory, every session begins from zero, and the temporary chat history that scrolls away does nothing to help. With it, you tell the agent once where things live and how you work, and it keeps that knowledge in a permanent place it can reach for on its own. The difference is the difference between an assistant who has been with you for years and a brand-new one every single morning who has never met you.
What is grounding?
Grounding is the practice of tying an agent’s answers to a real source instead of letting it answer from memory alone. AI models have a well-known failure: when they do not know something, they will sometimes produce a confident, plausible, and entirely invented answer. Grounding is the correction. You connect the agent to the actual material, your real documents, live data, verified records, and instruct it to answer from that. Ask a grounded agent about your refund policy and it reads the policy; ask an ungrounded one and it may simply guess. This is the most dependable lever there is for reducing those invented answers, which the field calls hallucinations, and it is why any serious business system keeps its agents anchored to real data.
What are guardrails?
Guardrails are the rules that define what an agent is and is not permitted to do: which tools it may touch, which are off limits, how far it can go before it must stop and ask. The need becomes obvious the moment real consequences enter the picture. You might happily let an agent draft and send invoices while never letting it move actual money. Guardrails are where you draw that line and others like it. As agents take on more responsibility, guardrails stop being a finishing touch and become the first thing you set up rather than the last.
What is sandboxing?
Sandboxing is running an agent in a safe, walled-off copy of the real environment, where nothing it does carries real consequences. Before you let a new agent operate on live accounts or genuine records, you let it work in the sandbox, a duplicate system in which a mistake costs nothing. The principle is the same as a flight simulator: you want the pilot to make every bad landing there, not on a runway with passengers aboard. When the price of an error is real money or lost data, the sandbox is where the errors are supposed to happen.
What is human-in-the-loop?
Human-in-the-loop means a person must approve certain actions before the agent carries them out. The agent does the work, then pauses at the consequential moments and waits for a sign-off. Routine, low-stakes actions flow through on their own; the ones that are expensive or hard to reverse, a large refund, a signed contract, a permanent deletion, stop and wait for a human yes. You decide where that line sits. The principle is simply that for the decisions that really matter, a person stays in the chair.
What is a context window?
A context window is the amount of information an agent can hold in mind at one time while it reasons. Picture the size of its working desk. Everything that fits on the desk is available to it at once; pile on more than fits, and the older material slides off the edge and out of view. This is why an agent can summarize a short document flawlessly yet start losing details from the opening chapters of a very long one. Context windows have grown remarkably large, but they remain finite, which is exactly why long-term memory matters, serving as a filing cabinet beside the desk for everything that will not fit on top of it.
What is a multi-agent system?
A multi-agent system is several agents working together as a coordinated team rather than one agent doing everything alone. An orchestrator directs the work, specialist subagents each handle their piece, shared memory keeps them all working from the same facts, and guardrails hold each one in its lane. It succeeds for the same reason a good team succeeds: divide the labor, match each task to the right specialist, and keep someone responsible for the whole. A single generalist agent attempting all of it would be slower and more error-prone, in just the way one overloaded employee would be.
How does it all fit together?
If any of the above blurred, here is the shape of it in one pass. An agent is software that does tasks rather than just chatting. It works by running a loop and using tools to take real actions. It connects to your other software through MCP, stays accurate by being grounded in your real data, and remembers what matters between sessions through memory. When a goal is large, an orchestrator splits it across specialist subagents working as a multi-agent team. And the whole operation is kept safe by guardrails on what is allowed, a sandbox for practice runs, and a human in the loop for the decisions that count, all of it bounded by how much the agent can consider at once, its context window.
| Term | What it is | Example |
|---|---|---|
| MCP | A shared standard for plugging AI into your tools | A CRM that “supports MCP” connects with no custom code |
| Agent loop | The perceive, plan, act, observe cycle an agent repeats | A booking agent keeps proposing times until one is accepted |
| Tool use | The ability to take real actions, not just describe them | It sends the email instead of drafting it for you to send |
| Orchestrator | The agent that breaks a goal into tasks and delegates | Turns “launch the promotion” into assignable pieces |
| Subagent | A specialist handling one narrow task | One agent writes copy while another only checks facts |
| Memory | Information that persists across sessions | The agent recalls your setup without being told again |
| Grounding | Tying answers to a real source to stay accurate | It quotes your actual refund policy rather than guessing |
| Guardrails | The rules for what an agent may and may not do | Allowed to send invoices, never to move money |
| Sandboxing | A safe copy of the real system for testing | Runs on fake records before touching live accounts |
| Human-in-the-loop | Required human approval for high-stakes actions | A person signs off before any large refund |
| Context window | How much an agent can hold in mind at once | It loses early details in a very long document |
| Multi-agent | A coordinated team of specialist agents | A manager agent plus several specialists, in sync |
You do not need all twelve to get value from any of this. Most real setups begin small, a loop, a couple of tools, one guardrail, and grow as the work demands. The vocabulary is worth knowing mainly so that when a vendor, an article, or a relative at dinner throws one of these words at you, you can tell which piece they mean, and whether it is a piece you actually need.
If you are trying to work out where any of this fits in your own business, which pieces matter and which are noise, that is the conversation we have every day. Let’s meet or call our AI voice agent at (530) 702-4447. [email protected].