Everyone talks about AI agents, MCPs, and RAGs like it's common knowledge. Half the room is pretending. Here's what those words actually mean:

LLM — a neural network trained on the whole internet that predicts the next word. Right, LLM doesn't "understand" things. LLM sounds confident even when it's wrong and hallucinates facts that don't exist. GPT, Claude, Gemini are all LLMs.

AI Agent — an LLM that takes actions, not just talk. A regular chatbot just answer your questions. An AI agent books a meeting, pulls data from your CRM, and drops your production database. Think crazy intern with API access.

Prompt — the instructions you give the model. Could be a one-liner, could be a 2-page doc with examples. The difference between "write me marketing copy" slop and getting something usable is usually the prompt. Prompt Engineering is a science of writing effective prompts.

MCP (Model Context Protocol) — a standard for connecting LLMs and AI agents to external tools and data sources. It gives AI structured access to outside world, like API integrations.

Skill — a reusable, packaged capability you hand to an agent. "Generate a PDF report from this data" or "deploy this branch to staging" as a self-contained unit the agent can pick up and run.

AI Harness — the code wrapped around an LLM that turns it into something useful. It handles the loop: send prompt, get response, execute tool calls, feed results back, repeat. The scaffolding between "raw model" and "AI agent".

Context Window — how much text the model can see at once. Small window means it forgets the beginning of your document before reaching the end. Bigger window fits more but costs more. 1M context window are the best we have today.

Token — the billing unit. Roughly ¾ of a word. Every API call charges for input tokens (your prompt) and output tokens (the response). A 10-page document is about 4,000 tokens. This is why "send the whole database" isn't a strategy.

RAG (Retrieval-Augmented Generation) — before answering, the system searches your documents, grabs the relevant chunks, and stuffs them into the prompt. Under the hood, text gets converted into numbers that capture meaning (embeddings) and stored in a vector database built for "find me something similar." This is how "chat with your docs" products work.

Inference Provider — a company that runs AI models so you don't have to buy GPUs. OpenAI, Anthropic, Google, Fireworks, Together. You send API calls, they send responses, you pay per token.

Human-in-the-loop — keeping a person in the decision chain. Most production agents use this for anything involving money or actions you can't undo. Remember, LLMs hallucinate and don't really understand things.

AGI (Artificial General Intelligence) — AI that can do anything a human can, across any domain. Doesn't exist yet. Timelines range from "2 years" to "tomorrow" depending on who you ask and what they're selling. Not useful for your Q3 planning either way.