Glossary
The Agent Advantage — every term, in plain language.
Search the list, or scan it. Definitions match what the book uses them to mean — not what the marketing decks want them to mean.
28 terms
Agent
A model wired into a loop with tools and a goal — a chatbot that got hands. It can browse, run code, edit files, and use software, and it can repeat that cycle unsupervised until a job is done. The employee, versus the encyclopedia.
Always-on agent
An agent that runs around the clock on its own machine (usually a rented server), doing scheduled work whether or not you're awake. Distinct from always-public — always-on means awake, not reachable by strangers.
API (application programming interface)
The pay-per-use way to access a model from your own software, billed by how much you send and receive. Contrast with a flat monthly subscription.
Blended cost
A single-number way to compare model prices, weighting input and output tokens by how you actually use them (this book's tool uses 85% input, 15% output). A fairer comparison than either number alone, except for reasoning models where output dominates.
Context window
A model's working memory — how much text it can hold "in mind" at once. Think of a desk of a fixed size; when it fills, older papers slide off the back. Explains about half of all agent confusion on long jobs.
Creative volume
In advertising, the modern lever that replaced targeting: since the ad platform's algorithm picks who sees what, your job is to feed it many good creative variations so it has winners to find. "Creative volume is the new targeting."
Frontier model
The current most-capable (and most expensive) tier of models — the ones to call only when a job genuinely needs the extra brain.
Hallucination
When a model states something false with confidence, because it doesn't know and isn't built to admit it. Manageable, not solved: reduced by giving it sources, self- and cross-checking, and a human review before anything publishes.
Local model
An open-weight model you run on your own hardware — no per-token bill, private, but limited by your machine's memory and speed. Great for text generation; slow for heavy reasoning (see memory bandwidth).
MCP (Model Context Protocol)
The open standard "plug" that connects models to tools — the USB-C of AI. You rarely think about it directly; you just notice that things connect.
Memory bandwidth
How fast a machine can move data between memory and processor — the real bottleneck for running models locally. Because generating each token means reading the whole active model once, a slow-bandwidth machine handles short text fine but crawls on long reasoning.
Model
The underlying "brain" that reads text and produces text — Claude, GPT, Gemini, and the open ones you can run at home. They come in sizes and prices; matching the size to the job is a core skill.
Moderation hold
A queue between an agent and the public where drafts wait for a human to approve, edit, or reject before anything publishes. The single architecture choice that prevents most public agent disasters.
MoE (Mixture-of-Experts)
A large model internally split into many "experts," only a few of which activate per token — so a huge model can run at the speed and memory cost of a much smaller one. Why some big models run locally when their full size shouldn't fit.
MoA (Mixture-of-Agents)
Several models (or passes) working a problem and a final one combining their answers — the cheap-drafters-plus-smart-judge pattern, formalized.
Open-weight model
A model whose weights are published so you can run it yourself. "Open-weight" is not automatically "free for commercial use" — the licenses vary and some carry conditions at scale (see Appendix C).
Orchestrator / worker bees
A cost pattern: run one smart, expensive model as the orchestrator to judge and direct, and many cheap models as workers to do the bulk drafting. Roughly flagship quality at a fraction of the price, because the costly brain only reads, it doesn't type.
Prompt
What you type to a model. Also prompt caching: paying a reduced rate to re-send the same unchanging chunk (like a long instruction set) instead of full price every time.
Quantization
Squeezing a model's parameters into fewer bits so it needs less memory to run. The common sweet spot (Q4) roughly halves the size for a small quality loss — the lever that makes local models viable on normal hardware.
Sub-agent
A temporary helper agent spun up to work part of a big job in parallel — like a foreman sending crews to different corners of a site.
Profile
A saved configuration that makes one agent behave as different specialists for different jobs — a meticulous researcher for one task, a punchy copywriter for the next — each with its own instructions and permissions.
Reasoning model
A model that "thinks" through thousands of hidden tokens before answering. More capable on hard problems, but far more expensive and slower than the token count suggests, because all that hidden thinking is billed and generated.
ROAS (return on ad spend) / CAC / LTV
The money-sanity numbers for advertising: what you get back per ad dollar (ROAS), what it costs to win a customer (CAC), and what a customer is worth over time (LTV). The one durable rule: keep CAC comfortably under LTV.
Structured output
A model returning clean, valid, machine-readable data (rather than prose) — a capability to verify, not assume, when an agent needs to hand data to other software.
Tool (in the agent sense)
Anything an agent can do beyond talking: search, run code, read a file, post to a store, send an email. An agent's usefulness rises with the quality of its tools.
Tool-calling
How reliably a model can actually use its tools — the capability that matters most when a model is driving an agent, more than raw intelligence.
Token
The unit models read and write in, roughly three-quarters of a word. Prices are quoted per million tokens, split between input (what you send) and output (what it writes).
VPS (virtual private server)
A small rented computer in a data center that never sleeps — the standard home for an always-on agent, for roughly the price of one coffee a month.