From LLM to Agent: 12 Core AI Concepts Explained in One Go

Overview

The article starts by noting the confusion caused by a flood of AI jargon such as agent, skill, script, MCP, and context window, and promises to clarify these terms through a single analogy that shows how a talking LLM can become an AI employee capable of doing work.

01 LLM: The Brain

LLM (Large Language Model) is described as the employee’s brain that only predicts the next token based on prior text. It can answer coding or design questions after extensive training, but it cannot perform actions on its own.

02 Tool: Hands and Feet

Tools are external functions that the LLM can invoke, turning textual output into actionable commands (e.g., “fetch this file”). The LLM still only outputs text, but the surrounding system executes the requested tool, enabling the AI to do things beyond speaking.

03 MCP: Universal Plug

MCP is a standardized interface (similar to a Type‑C connector) introduced by Anthropic in November 2024 that allows new tools to be attached to the LLM. Products that support MCP can install any compliant tool, but the plug itself contains no functionality; users must add tools themselves.

04 Script: On‑the‑Fly Programs

When existing tools are insufficient, the AI can write a short script—a piece of code executed by an existing tool—to accomplish the task (e.g., renaming 500 files). The script is likened to a 3D printer that creates a custom part when needed.

05 Prompt: The Work Order

A prompt is the instruction given to the AI employee. Clear prompts must specify what to do, what not to change, and how to determine completion, otherwise the AI may modify unintended parts.

06 Skill: Written SOP

Skill refers to a documented standard operating procedure that captures repeated instructions. The AI reads the SOP before acting, reducing the need to repeat the same prompt each time.

07 Token: Text Chunks

Tokens are the smallest text units the model processes. Tokenization balances speed and robustness; fewer tokens mean lower cost (“paying” fewer tokens) and faster processing.

08 Context Window: Desk Size

The context window limits how many tokens the model can consider at once, analogous to the size of a desk that can hold only a limited amount of material. Older information is pushed out as new content arrives.

09 RAG: Retrieval Assistant

When the needed information exceeds the context window, Retrieval‑Augmented Generation (RAG) stores documents in small chunks and retrieves the most relevant pieces for the model, preventing context overflow.

10 Loop: Self‑Driving Execution

Loop enables the AI to run a script, observe results, and decide the next step without continual prompting, effectively allowing the agent to operate autonomously.

11 Harness: The Office

Harness is the surrounding framework that bundles tools, permissions, context limits, and retrieval mechanisms. Different harnesses (e.g., Claude Code, Cursor) provide varying levels of convenience and productivity.

Conclusion

Combining these eleven components forms a functional agent—a LLM embedded in a system that can act, not just talk. Understanding the terminology lets readers decode future AI discussions.