Agentic AI Glossary: The Vocabulary You Need in a Meeting

This is the dictionary you need so you do not get lost when a vendor or your own team starts throwing terms around in a meeting about agents. Each entry is defined in two or three sentences, with nothing taken for granted, and when a concept has an article that develops it on this site, I link to it so you can go deeper. It is ordered alphabetically to be used as a reference, not read straight through.

If you are starting from zero, begin with the pillar article on AI agents for business, which gives the overall map, and this glossary is the companion for the fine print.

Terms, A to Z

Agent (AI agent). Software that takes an objective, decides the steps to meet it, and executes them against real systems, escalating to a person when the case exceeds it. It differs from a chatbot in that it acts, not just converses. It is the central concept of this whole vocabulary, and the detail is in the pillar article.

Architecture (agent). The full design of the system: how the model, tools, memory, orchestration and controls connect so an agent works reliably. It is the difference between a demo and something that survives production. I develop it in AI agent architecture.

Chatbot. A conversational interface that answers questions, whether through scripts and intents or a language model. It converses, but it does not execute work on your systems. The distinction from an agent is in chatbot or AI agent.

Context window. The amount of information a model can hold at once while processing a request: instructions, case data, history. It is limited, and managing it well (what goes in and what stays out) is part of the craft of building reliable agents.

Copilot. A tool that assists a person (suggesting, drafting, summarizing) while the human decides and executes. The productivity it delivers is individual, not the process’s. When a copilot is enough and when you need an agent is in copilot or AI agent.

Evals (evaluations). A set of tests with known answers that measure an agent’s quality before and during production. Without evals there is no serious way to claim an agent works. The method is in AI agent evaluation.

FDE (forward-deployed engineer). An engineer who deploys alongside the client to connect the system with their real data, rules and exceptions. They are the ones who translate a specific operation into agents that work. What they do exactly is in what is a forward-deployed engineer.

Governance (agent). The frame of rules, limits, permissions and responsibilities that lets an agent operate without gambling. It answers who decides what an agent may do and who answers for its actions. The full frame is in AI agent governance.

Hallucination. A model response that sounds plausible but is false or invented. In an agent that acts on systems, a hallucination is not a cosmetic error: it can translate into a wrong action, which is why evaluations and traceability matter so much.

Harness. The software scaffolding around the model that gives it what it needs to work: access to tools, context management, control of the decision loop and its limits. The model is the engine and the harness is the rest of the vehicle, where much of the reliability lives.

Human in the loop. An arrangement where a person validates or authorizes certain agent actions before they take effect. It is one of the levers for governing risk, and where you place it is a business decision that sits at the heart of AI agent governance.

LLM (language model). The model that understands and generates text and serves as the agent’s reasoning engine. It is a utility, contracted through an API and swapped without rebuilding the process. The options for business are mapped in LLMs for business.

MCP (Model Context Protocol). A standard for connecting a model to external tools and data sources uniformly, instead of hand-coding each integration. It lowers the cost of giving an agent access to systems. I explain it in MCP model context protocol.

Memory. An agent’s ability to retain information across steps or sessions: what it has done, what it knows about the case, what it learned from previous interactions. It differs from context in that it persists beyond a single request.

Model-agnostic. A property of a system that is not tied to one model provider and can swap the model for another depending on the task. It treats models as a utility, chosen by quality, cost and latency. How to decide which model goes on each task is in choosing an LLM for each task.

Multi-agent. A configuration where several specialized agents collaborate on the same job, each with its role, coordinated by an orchestration layer. It is used when a single agent falls short for the complexity of the process.

Orchestration. The coordination of the steps, tools and, where relevant, the several agents in a flow, deciding what runs, in what order and under what rules. It is the operation’s conductor. I develop it in AI agent orchestration.

Permissions and controls. The set of rules defining what an agent may read, may write and must never touch in each system. It is the concrete implementation of governance. How they are designed is in AI agent permissions and controls.

Prompt. The instruction given to the model, from the user’s request to the background rules that define its behavior. In an agent, much of the business logic is encoded here and in the tools it has available.

Prompt injection. An attack that tries to make the agent ignore its instructions by slipping malicious orders into the text it processes, for example inside an email or a document. It is one of the specific risks of agents that act on external data. The defenses are in AI agent security.

RAG (retrieval-augmented generation). A technique that gives the model access to your company’s information by retrieving it at answer time, instead of relying only on what the model ships with. It is how an agent connects to your knowledge. I cover it in enterprise RAG.

Risk (agent). The set of things that can go wrong when an agent acts on your systems: wrong actions, data leaks, untraced decisions. It is governed with controls, not by trusting the model to be good. The full taxonomy is in AI risks for business.

ROI. The return of an agent deployment, measured on the work it absorbs: cost per unit, volume processed, quality. How to build that measurement for a board is in AI operations ROI.

RPA (robotic process automation). An earlier technology that automates tasks by replaying recorded clicks and steps, with no judgment. It works with very stable processes and breaks on the first exception, which is exactly where an agent adds value.

Service as software. A model where the software does the work in the background and the client buys the service and its outcome, not a license to operate. It is the model we work under at Arkatai. I explain it in service as software.

Skill. A concrete, reusable capability given to an agent for a specific task, packaged so it can be invoked when needed. It lets you compose agents from pieces instead of programming each behavior from scratch.

Tool use. The agent’s ability to invoke external tools (query a database, call an API, generate a document) to act on the real world. It is what turns a model that only talks into an agent that does.

Trace. The record of what an agent did on each case: what it received, which rule it applied, which action it took, why it escalated. Without a trace there is no audit. How it is built and reviewed is in auditing AI agent decisions.

Frequently Asked Questions

What is agentic AI in a few words?

It is AI that does not just generate text or images but executes work: it takes an objective, decides the steps and acts on real systems, within defined limits. The piece that does that work is the agent, and the rest of the terms in this glossary describe its components and its controls.

Do I need to understand all this vocabulary to buy agents?

Not by heart, but enough to ask the right questions. The terms that will serve you most in a buying decision are governance, permissions, evals, trace and model-agnostic, which are the ones that separate a serious system from a pretty demo.

What is the difference between an agent, a copilot and a chatbot?

The chatbot converses, the copilot assists a person who does the work, and the agent executes the work on its own within permissions. Each boundary is developed in copilot or AI agent and chatbot or AI agent.

Why do evals and traces matter so much?

Because they are the two pieces that let you claim, with evidence rather than faith, that an agent works and is accountable. Evals measure quality before and during production, and the trace leaves an auditable record of every action. A vendor who cannot show you either is asking for blind trust.