Glossary
The Agentic AI Glossary
Agentic AI terms don't all have a stable definition yet. This glossary explains them one by one, drawing on sources that most people agree on.
This glossary gathers the terms that come up most often when talking about AI agents: autonomous agent, multi-agent orchestration, MCP, RAG, tool calling, BYOK, human-in-the-loop, observability, guardrails and governance. Some come from academic research, others from the technical documentation of major model providers, and others still from recent regulatory frameworks. Vocabulary in this field moves fast, and the same words sometimes cover fairly different things from one provider to the next.
Every page starts from the same principle: check what the reference sources actually say first (official documentation, a foundational research paper, a regulatory text), flag it when two serious sources don't quite agree, then explain the term simply, with a concrete example rather than an abstract definition. Where relevant, the example draws on how an autonomous AI agent works on Atako, without inventing anything that isn't documented.
The goal isn't to replace a dictionary definition, but to give you enough to understand a term quickly enough to keep reading elsewhere, whether that's a technical article, product docs, or a conversation with a team. Each page also links to related terms, so you can move from one concept to the next without losing the thread.
All terms
Agent Orchestration: Definition, How It Works, and Use Cases
Agent orchestration refers to coordinating several AI agents or sub-agents within a single system: task allocation, step sequencing, execution oversight, and result aggregation. An orchestrator, centralized or distributed, decides which agent acts, when, and how results get combined.
Agentic AI: Definition, How It Works, and Examples
Agentic AI is the paradigm of AI systems that perceive their environment, plan a sequence of actions, and execute them using tools, working toward a goal across multiple steps rather than simply answering a single isolated request once, with limited but adjustable human oversight.
AI Agent Observability: Seeing What an Agent Does in Real Time
Agent observability is the ability to track an AI agent's activity in detail: its tool calls, decisions, and errors, along with their origin and outcome, typically through logs, an event timeline, or traces. It makes it possible to understand why an agent acted a certain way and to catch a problem before it gets worse.
AI Agent: Definition, How It Works, and Examples
An AI agent is a program that perceives its environment, decides on an action, and carries it out, alone or via external tools, to reach a goal. It differs from a classic chatbot in its ability to chain several steps together and act, with a level of autonomy that varies by platform.
Autonomous AI Agent: Definition and Difference from a Workflow
An autonomous AI agent pursues a long-term goal, chooses its own sequence of actions to get there, and runs continuously in its own environment rather than being woken up for a single request. It only involves a human for a specific decision, unlike a workflow triggered on a one-off basis by an event.
BYOK: Running an AI Agent with Your Own API Key
BYOK (Bring Your Own Key) is an option that lets you run an AI agent or tool with your personal API key from a model provider (OpenAI, Anthropic, Mistral AI), rather than with the access included in the subscription. Model billing then goes directly through the provider's account, outside the platform's plan.
Generative AI vs. Agentic AI: What's the Difference?
Generative AI produces content (text, image, code) in response to a one-off request. Agentic AI plans a sequence of actions and executes them with tools to reach a goal across multiple steps. The latter generally uses the former as its internal reasoning engine.
Guardrails and Governance for AI Agents: Definition and Mechanisms
Guardrails and governance refer to the full set of rules, permissions, and controls that define what an AI agent is allowed to do, before, during, and after an action executes. This covers per-action permissions, human validation on sensitive decisions, auditing of what happened, and the ability to cut off access at any time.
Human-in-the-Loop: Keeping a Human in an AI Agent's Loop
Human-in-the-loop is a design principle where a person retains the authority to approve, correct, or block a decision or action generated by an AI system, at a specific point in the process, before it produces a real effect. It's a control mechanism, not continuous oversight of every step.
MCP (Model Context Protocol): Definition, How It Works, and Examples
MCP (Model Context Protocol) is an open protocol, created by Anthropic in late 2024, that standardizes how a model or AI agent connects to external data sources and tools. It replaces one-off custom integrations, built per tool, with a common language between AI applications and third-party systems.
Multi-Agent System: Definition, How It Works, and Examples
A multi-agent system brings together several autonomous AI agents that work together on the same complex task. Each agent has a specialized role, exchanges information with the others, and makes its own decisions. This division of labor makes it possible to handle workflows a single agent couldn't manage efficiently on its own.
RAG (Retrieval-Augmented Generation): Definition and How It Works
RAG (Retrieval-Augmented Generation) is a method that pairs a document search engine with a language model: before answering, the model retrieves relevant passages from an external base, then generates its response from those passages. This limits fabrication and allows the use of recent or company-internal information.
Tool Calling: How an AI Agent Calls External Tools
Tool calling (also called function calling) is a language model's ability to identify that a request requires an external action, such as reading a database or sending a message, and to produce a structured call request with its arguments. An application then executes that call and returns the result to the model.
Frequently asked questions
What is an AI agent, in one sentence?
An AI agent is a system that runs continuously, makes decisions, and can act on real tools to reach a goal, rather than just answering a question in a one-off conversation. The site's pillar page details this definition and what sets an autonomous agent apart from a plain conversational assistant.
What's the difference between an AI agent and classic generative AI?
Classic generative AI answers a one-off request with no memory and no ability to act beyond the text it produces. An AI agent keeps memory across exchanges, can call external tools, and keeps working after its first response, until the goal is reached or a human steps in.
Is this glossary updated regularly?
Yes. Each term shows a last-updated date and draws on reference sources cited at the bottom of the page, such as official documentation from model providers or recognized AI governance frameworks. Vocabulary in this field moves fast, and definitions are revised accordingly.
What to read next
CTO at Atako
This content was written by Atako's AI agents, then reviewed, corrected, and approved by Romain Laodicina, CTO of Atako.