AI Agent Explained: From Models and Tools to Skills and Harness Engineering

What an AI Agent Actually Is

An AI Agent is a system whose core is a large language model (LLM) that can invoke external tools, receive feedback, and iteratively complete a task. Unlike a single‑turn chat model that answers one question, an Agent receives a goal, reasons about the next step, executes it, and repeats based on the result.

AI Agent 是一个以大模型为核心、能够调用工具、接收反馈并持续完成任务的系统。

Model vs. Agent

The model itself is a pure "text‑in, text‑out" function without cross‑call memory or an execution loop. It can express the intent to use a tool, but the actual tool call, web interaction, file access, or API request is performed by the surrounding system, often called the harness.

Scaffolding and Harness

These two components are frequently grouped as an "Agent framework" but serve distinct purposes.

Scaffolding 管“怎么想”，Harness 管“怎么跑”。

Scaffolding decides *how to think*: prompts, context, and reasoning. Harness decides *how to run*: the execution loop, tool routing, and state management.

Context Engineering and Policy

Context Engineering defines exactly what the model sees at each step—system prompts, tool descriptions, conversation history, retrieved knowledge, and tool results. The context is updated continuously as the task progresses.

Policy defines the behavior the Agent follows when choosing among possible actions. In reinforcement‑learning terms it is a probability distribution over actions; in LLM agents part of the policy is baked into the model weights, while the rest is shaped by prompts, tools, memory, and the execution loop.

Tool, Skill, Sub‑Agent

Tool is the most basic layer: an external capability the Agent can reach, such as an API, code interpreter, database, web search, or file system. The model only expresses the intent to use a tool; the harness actually performs the call.

Skill is a reusable method—a set of steps, experience, and a relatively stable workflow for achieving a goal (e.g., bug investigation, data cleaning, market‑research summary). It represents the Agent’s “routine”.

Sub‑Agent is an independent Agent that can think, call tools, and complete a sub‑task on its own. For example, a main Agent tasked with writing an industry analysis may delegate data collection, data cleaning, and draft writing to separate sub‑agents before merging the results.

Training‑Stage Terminology

Environment is the space the Agent interacts with—browsers, file systems, code repositories, or any abstract task space.

Rollout records a complete execution from start to finish, capturing what the Agent saw, what actions it took, and the final outcome.

Reward scores the result of a rollout, indicating how well the Agent performed. Rewards can come from pass/fail tests, human preferences, or other evaluation metrics.

Trainer uses many rollouts and their rewards to update the model, enabling the Agent to improve its policy through repeated trial and error.

Illustrative Scenarios

Search for information and synthesize a summary.

Read a file and analyze its contents.

Invoke a code tool to process data.

Perform a sequence of actions on a web page.

These multi‑step tasks differentiate an Agent from a plain chat model.

Why Training Discusses Environment, Rollout, Reward, and Trainer

During the construction phase the focus is on how to assemble the system (model, tools, harness). In the training phase the focus shifts to improving the Agent’s ability to act in its environment. The environment provides the context for actions; a rollout captures a full episode; the reward evaluates the episode; and the trainer updates the model based on many rollouts and rewards.

Concept Diagram (textual)

AI Agent = Model + Tools + Harness
Scaffolding (how to think) → prompts, context, reasoning
Harness (how to run) → execution loop, tool routing, state
Context Engineering → what the model sees each step
Policy → action‑selection behavior
Tool → external capability (hand)
Skill → reusable routine (method)
Sub‑Agent → independent mini‑Agent
Environment ↔ Agent actions ↔ new state
Rollout → full episode
Reward → performance score
Trainer → updates model from rollouts

Final Takeaway

An AI Agent is not a new model name; it is an entire system built around a model. The model handles understanding and decision‑making, tools provide the means to act, and the harness drives the task forward step by step. Distinguishing these concepts clarifies the rapidly growing ecosystem of Agent products, frameworks, and research.