FOWL AI · Glossary

The AI terms everyone assumes you know.

Five words showing up in job postings, eval projects, and every AI conversation this year — defined in one plain sentence each, then tied to what they actually mean for your work.

Almost nobody explains these terms in order, from the ground up, without assuming you already know the last one — so they stack: you half-understand "LLM," which makes "Agent" fuzzy, which makes "MCP" sound like alphabet soup. This page fixes that, and stays updated as new terms enter the AI job market.

01 · Foundation
LLM
Large Language Model

Software trained on enormous amounts of text to predict the next word — which turns out to be enough to hold a conversation, write code, summarize a document, or answer a question.

An LLM doesn't "know" facts the way a database does. It's learned patterns from the text it was trained on and generates a response one piece at a time, each piece chosen based on everything that came before it. ChatGPT, Claude, and Gemini are all products built around an LLM — the LLM is the engine, the chat app is the car built around it.

Where you'll see this — Job postings and eval projects that mention "LLM evaluation" or "model output review" are asking you to judge how well one of these models answers, reasons, or follows instructions, not to build the model itself.
02 · Built on an LLM
Agent
AI Agent

An LLM that's been given a goal, a set of tools it's allowed to use, and the ability to keep working in steps until the goal is done — instead of answering one question and stopping.

A plain chatbot answers what you type and waits. An agent can be told to complete a task — research a topic, fix a bug, book something — then actually take the steps to do it, checking its own work and adjusting along the way, without a human approving every single move.

Where you'll see this — "AI Agent Orchestration," "agentic evaluation," and "multi-step task testing" roles on platforms like Mercor and Outlier all pay people to judge how reliably an agent completes real, multi-step work.
03 · Connects agents to your tools
MCP
Model Context Protocol

A standard, created by Anthropic and now adopted industry-wide, that lets an AI model connect to outside tools and data — your files, a database, a calendar — without a custom integration built for every single pairing.

Before MCP, connecting an AI to your calendar or a company's internal systems meant a one-off integration, built by hand, every time. MCP works more like a universal power outlet: any tool that speaks MCP can plug into any AI model that supports it, and they understand each other automatically.

Where you'll see this — If a project asks you to "test an MCP server" or "evaluate tool use," they mean checking whether a model correctly uses one of these plug-in-style connections, not writing the connection yourself.
04 · Teaches an agent a task
Skills
Agent Skills

A packaged, reusable set of instructions — usually a folder with an instructions file and any supporting resources — that teaches an AI agent how to do a specific task consistently, so it doesn't have to be re-explained every time.

Instead of typing out the same detailed instructions every time you want an agent to follow a specific checklist or writing style, you write it once as a Skill. The agent loads it by name whenever the task calls for it, the same way a new hire follows a documented process instead of guessing.

Where you'll see this — Newer eval projects ask evaluators to test whether an agent picks the right Skill for a task, or to author a new one — this is starting to show up in listings as "Skill authoring" or "Agent Skill design."
05 · An agent in your terminal
Claude CLI
Claude Code

Anthropic's command-line tool that runs Claude as a coding agent directly in your terminal — reading files, running commands, and editing real code across a whole project to complete a task, not just chatting about it.

Where a chat window gives you an answer, Claude Code actually does the work inside your codebase: it can read your files, make edits, run your tests, and iterate until a task is done. It's one of the clearest everyday examples of an "agent" (see above) actually at work.

Where you'll see this — Listings that mention "Claude Code eval," "agentic coding," or "CLI-based coding tasks" are paying you to work inside this tool (or a close equivalent) and judge how well it completes real engineering work.

New terms land in every issue.

FOWL AI breaks down what's actually changing in AI jobs and tools — every Monday, free.

Subscribe free →