Stop Misunderstanding AI Agents: A Clear Guide to All Core Concepts

Agent definition

AI Agent is a system that places a large model at its core, can call tools, receive feedback, and continuously complete tasks.

The key distinction is “continuously complete tasks” rather than a single turn of text generation. A regular chat model answers one question per turn, while an Agent receives a goal, iteratively decides the next step, executes it, observes the result, and repeats until the goal is achieved.

Typical multi‑step tasks include:

Searching for information and summarizing it

Reading a file and analyzing its contents

Calling a code tool to process data

Performing a sequence of actions on a web page

Model vs. Agent

The model (large language model) is the core of an Agent but does not constitute the whole system. A model is essentially “text in, text out”; it lacks persistent memory and cannot execute loops by itself.

The model can express an intention such as “I want to call a tool”, but the actual tool invocation, web interaction, file access, or API call is performed by the surrounding system (the harness).

Scaffolding and Harness

Scaffolding handles “how to think”; Harness handles “how to run”.

Scaffolding defines the reasoning process (what the model should consider at each step). Harness implements the execution loop that calls tools, manages state, and feeds results back to the model.

Context Engineering and Policy

Context Engineering

Context Engineering determines what the model sees at each step of an Agent’s execution. It includes system prompts, tool descriptions, conversation history, retrieved knowledge, and tool results. The harness continuously decides which information to keep, discard, or re‑inject as the task progresses.

It applies to both training and inference. During training, an incorrect context can bias the model; during inference, a wrong context can often be corrected by adjusting prompts or re‑injecting information.

Policy

Policy defines the behavior of an Agent: given a situation, how it chooses among possible actions. In reinforcement learning this is a probability distribution over actions; in LLM Agents the policy is partly encoded in model weights and partly shaped by prompts, tools, memory, and the execution loop. Policy is not the Agent itself; the Agent is the full system that acts in an environment, while Policy is the decision‑making strategy it follows.

Tool, Skill, Sub‑agent

Tool

A Tool is a concrete action that extends the Agent beyond its own code, such as calling an API, using a code interpreter, accessing a database, performing web search, or manipulating the file system. The model only expresses the intent to use a tool; the harness routes the call, obtains the result, and continues the loop.

Skill

A Skill is a reusable method—a set of steps, experience, and stable workflow—to achieve a specific goal (e.g., debugging a bug, cleaning data, writing a market‑research summary). It represents the “routine” of the Agent.

Sub‑agent

A 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 “write an industry analysis” can delegate to Sub‑agents that collect data, organize it, and draft sections, then aggregate the results.

Training‑related concepts

Environment

Environment is the space with which an Agent interacts—browsers, file systems, code repositories, or abstract task spaces. The Agent takes actions, and the Environment returns new states and results.

Rollout

A Rollout is a complete execution of a task from start to finish, recording what the Agent saw, what actions it took, and the final outcome.

Reward

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

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

Conclusion

AI Agent is not a single new model name; it is an entire system built around a model: the model handles understanding and decision‑making, tools provide actions, and the execution system (harness) drives the task forward step by step.

Code example

本文
约2200字
，建议阅读
5
分钟
本文介绍了 AI Agent 核心定义，厘清行业各类易混淆的专业术语与逻辑。