The article clarifies AI agent terminology that is often used inconsistently.

It centers on the distinction between a model (the LLM, text in/out, stateless) and the surrounding layers.

Scaffolding is the behavior-defining layer: system prompt, tool descriptions, format parsing, and context management.

Harness is the execution layer that calls the model, handles tool calls, and decides when to stop. Together, Agent = Model + Harness.

Products like Claude Code and Codex are specific harnesses built on specific models.

Context Engineering covers designing what goes into the context window (prompts, history, retrieved knowledge) and managing memory.

Policy is the behavior an agent follows, partly learned in weights, partly shaped by scaffolding and harness.

Tool use lets the agent act via APIs, code interpreters, etc.

Skills are reusable multi-step knowledge packages; sub-agents are independent agents called by another agent.

The training section defines RL Environment (stateful interaction), Trainer (runs episodes and updates weights), Rollout (full agent trajectory), and Reward (score for learning, with types like verifiable, learned, sparse, dense).

The article is a practical glossary, not a universal standard, aiming to reduce confusion in the rapidly evolving agent field.