第四篇 · PART 4

主流 AI 编程工具 Popular AI Coding Tools

从 IDE 补全到桌面智能体,2026 年的"数字同事"已经远超"写代码"本身——编程只是它们能力的起点。 From IDE autocomplete to desktop agent workbenches, 2026's "digital colleagues" go far beyond writing code — coding is just the entry point.

从编程 Agent 到通用 Agent From Coding Agents to General Agents

2024–2025 年,编程 Agent 只做一件事:写代码。到了 2026 年,几乎每个主流编程工具都在外溢成通用 Agent:能读文件、管浏览器、操作桌面应用、处理数据与文档。选工具时别再只问"它写代码强不强",而要问"它能不能成为你的通用数字同事"。

In 2024–2025, coding agents did one thing: write code. By 2026, nearly every mainstream coding tool has spilled over into a general-purpose agent: reading files, driving browsers, operating desktop apps, processing data and documents. Don't just ask "is it good at code?" — ask "can it be my general digital colleague?"

🧩

腾讯 WorkBuddy Tencent WorkBuddy

定位是桌面智能体工作台而非纯代码工具:接管桌面操作、文件与浏览器,被评价为"更适合国人体质的 Codex"。

Positioned as a desktop agent workbench, not a pure coding tool: drives desktop apps, files and browser — "a Codex built for Chinese users".

🛰️

阿里 Qoder Alibaba Qoder

由"通义灵码"进化而来:从代码助手扩展到研发全流程与办公场景,个人社区版免费。

Evolved from Tongyi Lingma: from code assistant to full dev-lifecycle and office scenarios; free community edition.

🌙

Kimi Code Moonshot Kimi Code

月之暗面出品,背靠"人人可用的通用智能助手"愿景:多模态编程工具 + 开源 CLI。

From Moonshot AI, backed by the vision of a general-purpose assistant for everyone: multimodal coding tool + open-source CLI.

⚙️

Claude Code / DeepSeek Harness Foreign & open-source signals

国外以 Claude Code 为代表,也早已是"通用任务 Agent";开源侧 DeepSeek Harness 直接定位通用 Agent 运行时。方向一致:编程是入口,通用是终点

Abroad, Claude Code already works as a general task agent; on the open side, DeepSeek Harness is a general agent runtime. Same direction: coding is the door, generality is the destination.

国内产品(推荐)Domestic Products — Recommended

🇨🇳 为什么主推国内 Why we recommend domestic first 访问快、支付方便、文档与支持是中文、合规友好;能力上与国外头部差距已经很小(详见"如何选择")。国外产品除非开源可自托管,否则在国内使用常有网络、支付与合规麻烦。 Fast access, local payment, Chinese docs, compliance-friendly; the capability gap vs foreign leaders has narrowed a lot. Unless a foreign product is open-source (self-hostable), it's usually friction to use from China.
工具 Tool 出品方 Vendor 形态 Form 开源 Open 收费模式 Pricing 核心优势 Strengths 适合谁 Best for
Qoder 阿里云 Alibaba Cloud IDE 助手 + Agent IDE assistant + agent 闭源 个人社区版免费;Credits 计费 free community edition; credit-based 研发全流程 + 办公场景,阿里生态 full dev lifecycle + office; Alibaba ecosystem 阿里云用户、想从代码走向全流程 Alibaba Cloud users
Trae 字节跳动 ByteDance AI 原生 IDE AI-native IDE 闭源 免费 free 开箱即用、中文友好、多模型 out-of-box, Chinese-friendly, multi-model 图形界面新手、快速上手 beginners wanting a GUI
Kimi Code 月之暗面 Moonshot AI CLI + IDE 集成 CLI + IDE integration CLI 开源 Kimi 订阅 / API 按量 Kimi subscription / API usage 多模态、通用助手定位、开源 CLI multimodal, general-assistant vision, open CLI 终端党 + 想要通用能力 terminal fans wanting generality
WorkBuddy 腾讯 Tencent 桌面智能体工作台 Desktop agent workbench 闭源 以官方为准 per official pricing 接管桌面/文件/浏览器,办公通用 drives desktop, files, browser; office-wide 想要"全能数字同事" those wanting an all-round colleague

国外产品(简述)Foreign Products — In Brief

以下产品值得了解(很多仍是能力标杆),但闭源产品在国内使用有网络/支付/合规门槛;开源部分可自托管、无支付障碍

Worth knowing (many are still capability benchmarks), but closed-source ones carry network/payment/compliance friction in China; open-source ones can be self-hosted without payment barriers.

工具 Tool 形态 Form 开源 Open 一句话定位 One-liner
Claude Code(Anthropic) 终端 CLI Terminal CLI 闭源 终端 Agent 标杆,自主执行长任务、支持 MCP the terminal-agent benchmark; long autonomous tasks, MCP
Cursor(Anysphere) AI 原生 IDE AI-native IDE 闭源 AI 原生 IDE 的代表,Tab 补全与多文件 Agent 体验顺滑 flagship AI-native IDE: Tab completion + multi-file agent
GitHub Copilot(Microsoft) IDE 助手 IDE assistant 闭源 普及度最高的 IDE 助手,与 GitHub 生态深度绑定 most widely used IDE assistant, deep GitHub integration
OpenCode 终端 CLI Terminal CLI 开源 MIT 开源 Claude Code 替代品,社区活跃、可自托管 open-source Claude Code alternative; active community, self-hostable
Codex CLI(OpenAI) 终端 CLI + 云端 Terminal CLI + cloud 开源 本地 + 云端沙箱双模式 local + cloud sandbox modes
Gemini CLI(Google) 终端 CLI Terminal CLI 开源 免费层 + 多模型切换,Google 账号即用 free tier, multi-model, Google account
📌 客观看待差距 A fair take on the gap 国外头部(尤其 Anthropic、OpenAI)在一些场景仍略先进;但差距已经很小——DeepSeek / Qwen / Kimi 的开源模型在多项基准上紧追甚至反超闭源。除中美之外,其他地区基本没有能跟上竞争的产品;对国内用户,国产 + 开源是性价比与合规兼得的主流选择。 Foreign leaders (esp. Anthropic, OpenAI) are still slightly ahead in some scenarios, but the gap is small — DeepSeek / Qwen / Kimi open models rival or beat closed ones on many benchmarks. Outside the US and China, almost nothing keeps pace. For users in China, domestic + open-source is the mainstream choice balancing cost, capability and compliance.

如何选择(国内视角)How to Choose (China Perspective)

  • 主力推荐(国内)
    • 开箱即用、图形界面Trae(字节,免费)want a GUI out of the box → Trae (free)
    • 需要研发全流程 + 办公、阿里生态 → Qoder(社区版免费)full dev lifecycle + office, Alibaba ecosystem → Qoder
    • 喜欢终端、可控、开源Kimi Code CLI(开源)terminal & open source → Kimi Code CLI
    • 想要全能数字同事(不只编程)→ WorkBuddy(腾讯桌面工作台)an all-round digital colleague → WorkBuddy
  • 国外产品:开源可自托管的(OpenCode、Codex CLI、Gemini CLI)随便用;闭源标杆(Claude Code、Cursor、Copilot)了解即可——除非团队确有需要并解决了网络与支付问题。Open-source foreign tools are fine anywhere; closed benchmarks are for reference unless you've solved access & payment.
  • 进阶玩法:让国产与国外工具互相 review 代码(Qoder 写、Claude Code 审),用"双保险"压缩幻觉。Pro tip: have domestic and foreign tools review each other's code for a second opinion.

通用工作流:与 AI 结对编程 A Universal Workflow for Pairing with AI

  1. 拆任务:把大目标拆成 15–60 分钟的小任务,一次只让 Agent 做一件事 Split: break big goals into small tasks; one thing per prompt
  2. 先探索再动手:让 Agent 先读文件、打印结构,再写处理逻辑 Explore first: let it read files and print structures before writing logic
  3. 小步验证:每次改动后运行,确认无误再继续 Verify small steps: run after each change
  4. 报错原样回贴:把终端输出整段贴给 Agent,比描述"它报错了"有效得多 Paste errors verbatim: far more effective than saying "it failed"
  5. 要求测试与文档:让 Agent 写单元测试、注释和 README Demand tests & docs
🙋 一个"差"的提问 A weak prompt
我的程序跑不了,帮我看看。
"My program doesn't run, help me."
🤖 一个"好"的提问(模板) A strong prompt (template)
任务:处理射电数据立方体,生成矩图。
数据:/data/hi_cube.fits,RA/DEC + 速度轴,单位 m/s,LSRK。
步骤:1) 用 astropy 打印 FITS header 关键键;2) 用 spectral-cube 加载并掩膜 NaN;3) 生成 3σ 掩膜的 moment 0/1/2;4) 保存结果并保留 WCS。
约束:内存有限,考虑 dask 后端;完成后打印统计信息。
请先给出计划,再逐步实现,每一步运行验证。
Task: process a radio data cube, produce moment maps. Data: /data/hi_cube.fits, RA/DEC + velocity axis, m/s, LSRK. Steps: 1) print key FITS header keys with astropy; 2) load with spectral-cube, mask NaNs; 3) make 3σ-masked moment 0/1/2; 4) save preserving WCS. Constraints: limited memory, consider dask; print stats at the end. Give a plan first, then implement step by step, verifying each step.

工作流本身也是一种"技能"——把常用流程沉淀成 Skill / 插件 / SOP,可以大幅提升效率。详见 技能生态篇

A workflow itself is a "skill" — turning recurring procedures into Skills, plugins or SOPs boosts efficiency a lot. See the Skills & Ecosystem page.

下一步:射电天文实战 → Next: Radio Astronomy in Practice