Agent The Autonomous Worker
模型是大脑，Harness 是身体与工具箱；而 Agent，是"带着目标、自主循环、直到把事情做完"的那套系统。 The model is the brain, the harness is the body and toolkit; the Agent is the system that, given a goal, loops autonomously until the job is done.
什么是 Agent What Is an Agent
AI Agent（智能体） = 模型（思考能力） + Harness（工具与运行环境） + 目标 + 自主循环（观察 → 思考 → 行动 → 再观察）。与"你问一句、它答一句"的聊天不同，Agent 会把任务拆成多步、亲自动手、根据结果调整方案，直到达成目标。
An AI Agent = model (thinking) + harness (tools & runtime) + a goal + an autonomous loop (observe → think → act → observe again). Unlike one-shot chat, an agent breaks the task into steps, does real work, adjusts based on results, and keeps going until the goal is met.
Model = how deep it can think; Harness = what it can safely do; Agent = whether it actually finishes the job.
核心机制：ReAct 循环 Core Mechanism: The ReAct Loop
ReAct（Reason + Act，推理 + 行动）是多数 Agent 的基本运转方式。下面的循环会自动播放，也可以点击任一步骤查看：
ReAct (Reason + Act) is how most agents operate. The loop below plays automatically; click any step to inspect it.
🔁 循环意味着什么 What the loop means
- 每一步都基于真实结果决策，而不是一次性猜答案 Each step decides on real results, not a single guess
- 出错可以自己发现、自己修复（跑代码 → 看报错 → 改代码）Can detect and fix its own errors (run → read error → fix)
- 长任务被拆成多步执行，可中途汇报进展 Long tasks are broken into steps with progress checkpoints
⚡ 与"一次问答"的区别 vs. one-shot chat
| 聊天 Chat | 问 → 答 → 结束 ask → answer → end |
| Agent | 目标 → 观察 → 行动 → 反馈 → … → 完成 goal → act → feedback → … → done |
Agent 的常见类型 Common Types of Agents
编程 Agent Coding Agents
写代码、改 bug、跑测试、重构。代表：Claude Code、OpenAI Codex CLI、Cursor Agent、Aider、Cline。
Write code, fix bugs, run tests, refactor. E.g., Claude Code, Codex CLI, Cursor Agent, Aider, Cline.
详见下篇 → see next page研究 / 浏览器 Agent Research / Browser Agents
自主上网检索、阅读、汇总成报告。代表：OpenAI Deep Research、Gemini Deep Research、Manus。
Autonomously browse, read, and synthesize reports. E.g., OpenAI Deep Research, Gemini Deep Research, Manus.
云端通用 Agent Cloud General-Purpose Agents
在云端沙箱里执行整套任务（软件工程、数据处理）。代表：Devin、Manus、OpenAI Codex 云端模式。
Execute whole jobs in cloud sandboxes (software engineering, data work). E.g., Devin, Manus, Codex cloud mode.
多 Agent 协作 Multi-Agent Orchestration
多个 Agent 分工（研究员 / 程序员 / 审查员），通过框架编排。代表：CrewAI、LangGraph，以及 Harness 里的子代理/工作流。
Multiple agents with roles (researcher/coder/reviewer) orchestrated by frameworks. E.g., CrewAI, LangGraph, and harness subagents/workflows.
风险与边界 Risks & Boundaries
- 权限失控：若没有沙箱，Agent 可能误删/误改系统文件 Runaway permissions: without a sandbox it could delete or modify system files
- 成本失控：循环可能无限重试，消耗大量 API 费用 Runaway cost: loops may retry endlessly, burning API budget
- 幻觉放大：一次编造可能被当成"真实结果"继续传播 Amplified hallucination: one fabrication can cascade as "real results"
- 死循环：目标无法达成时可能原地打转 Infinite loops: may spin when the goal is unreachable
—— 这正是 Harness 的沙箱、审批、资源上限与评测存在的意义：给自主加上"安全带"。
— Exactly why the harness's sandbox, approvals, resource caps, and evaluation exist: seatbelts for autonomy.
小结与下一站 Summary & Next Stop
Harness
身体 + 工具箱 + 安全规程，让模型能动手。
Body + toolkit + safety rules; gives the model hands.
回顾 Review →Agent
带着目标、自主循环、把事做完的系统。
The system that pursues goals autonomously.
下一站：编程 Agent 大观 → Next: coding agents