Jorge Laurel
A

Appendix · Agentic AI, Explained

Glossary & Index

Plain-language definitions of every recurring term, plus an index pointing at chapters

Plain-language definitions of the recurring terms used throughout this book, in the same practical spirit as the chapters themselves.

A2A (Agent-to-Agent protocol)
A communication standard that lets independent agents, built by different companies or on different frameworks, hand off tasks and status to one another.

Agent (AI agent)
A software system that can pursue a goal by making decisions, interacting with its environment, and taking actions, often using a language model and external tools.

Agent harness (runtime)
The surrounding software that decides what tools an agent may access, executes its tool requests, and reports results back to the model.

Agent washing
Marketing a product as an “AI agent” when it actually runs a fixed script rather than deciding its own steps.

Autonomy
The degree to which an agent can make decisions and carry out multiple steps toward a goal without human direction at every stage.

AutoGen / AG2
Microsoft’s multi-agent framework, recently rearchitected and rebranded AG2, built around multiple agents debating and revising an approach together.

Chatbot
A conversational AI system designed primarily around dialogue. Some modern chatbots can also use tools or take actions, so the boundary between chatbot and agent is not absolute.

Coding agent
An agent purpose-built for software development that can read, write, and test code, and adjust based on the results.

CrewAI
A role-based framework for building multi-agent systems, where agents collaborate the way members of a small team would.

Function calling (tool calling)
The mechanism by which a model outputs a structured request to run a specific function, which surrounding software then executes and reports back on.

Graph-based orchestration
A way of structuring a multi-agent system as an explicit flowchart of steps and branches, rather than as role-playing team members.

Hallucination (confabulation)
When a model states something false with the same fluent confidence it uses for something true, because it is generating a plausible sequence of text, not recalling a verified fact.

Human-in-the-loop
A design practice where a person reviews or approves an agent’s progress at set checkpoints, rather than letting it run entirely unsupervised.

Instructions (system prompt)
The text that defines an agent’s goal, boundaries, and expected tone or format.

Knowledge cutoff
The point in time after which information is generally not included in a model’s training data. External tools and retrieval systems can give the model access to newer information.

LangGraph
A graph-based framework for orchestrating agents as an explicit, debuggable flowchart of steps, popular for tighter control over long-running tasks.

Large language model (LLM)
A statistical system trained to predict the next word in a sequence of text; the foundation underneath most modern AI products, including agents.

Memory
Information or state retained so an AI system can use it again later in a task or future interaction. Memory may be stored outside the model, within a conversation context, or through other state-management mechanisms.

Model Context Protocol (MCP)
An open standard, introduced by Anthropic in November 2024, for connecting AI applications with compatible tools and data sources through a consistent protocol instead of requiring a separate integration pattern for every pairing.

Multi-agent systems
An architecture in which several specialized agents divide a complex task among themselves, the way a human team divides labor.

n8n
A visual workflow-automation platform that lets people wire together triggers, steps, and outputs, including AI-enabled and agentic workflows, without writing a framework-style codebase.

Prompt
The text a person provides to a model, which becomes the starting point for its next-word predictions.

Prompt engineering
The practice of writing specific, well-structured prompts to guide a model toward a more useful output.

Prompt injection
An attack in which instructions hidden in content an agent reads, a webpage, file, or email, are followed as if they were legitimate commands from the user.

ReAct (Reason + Act)
A research approach introduced in a 2022 paper that interleaves reasoning and action, allowing a language model to use observations from its actions to update what it does next.

Role-based teams
A way of structuring a multi-agent system where each agent is assigned a role, similar to members of a human team.

Think-act-observe loop
The repeating cycle at the center of agent behavior: the model decides a next step, a tool carries it out, the result is observed, and the cycle repeats.

Tool
A defined function an agent is allowed to call, such as searching the web, reading a file, or sending a message.

Training
The resource-intensive process of training a model on large datasets so that its parameters learn useful patterns. Models can subsequently be fine-tuned, post-trained, updated, or replaced with newer versions.

Index

Terms below link to every chapter that discusses them (markdown has no page numbers, so this index points to chapters instead).

A

C

F

G

  • Graph-based orchestration — Ch. 6

H

  • Hallucination (confabulation) — Ch. 1
  • Human-in-the-loop — Ch. 3

I

K

  • Knowledge cutoff — Ch. 1

L

M

N

P

R

  • ReAct (Reason + Act) — Ch. 4
  • Role-based teams — Ch. 6

T