Google Antigravity 是什麼?

它是 Google 推出的 agent 優先開發平台,最早於 2025 年 11 月跟 Gemini 3 一起亮相,2026 年 I/O 升級為 2.0,變成圍繞多 agent 編排的獨立桌面應用,並新增 CLI、SDK 與企業支援。

Antigravity 2.0 跟一般 AI 寫程式工具差在哪?

傳統工具是 AI 在你打字時補全;Antigravity 是 agent 優先,你交付任務、AI agent 自己跨編輯器、終端機、瀏覽器完成,人退到設計任務與審查驗收的角色。

它會讓工程師失業嗎?

短期不會。agent 會出錯、可能誤動檔案,團隊需要的是把人力從打字轉到設計任務與嚴格審查。工程師的價值會從『寫得快』位移到『審得準、設計得好』。

台灣開發者該怎麼開始?

建議先用個人專案、非正式環境試跑 agent,觀察它在哪會出包,累積『帶 agent』的經驗。導入正式專案前務必做好版本控制與權限隔離。

Google Antigravity 2.0 上線:當寫程式的主角從「你」變成「AI agent」

Google 在 2026 年 I/O 把 Antigravity 升級到 2.0,從一個 IDE 變成獨立的桌面 agent 平台,還附帶 CLI、SDK,甚至把原本的 Gemini CLI 整併進來。對天天寫程式的人來說,這代表一件事:AI 不再只是幫你補全,而是準備替你「把整段活幹完」。

At 1 a.m., a backend engineer in Hsinchu who hadn't gone to sleep yet posted a screenshot in a group chat: on the screen was not the familiar code editor, but a row of AI agents "running tasks on their own." He commented, "In the future, will I just need to review code?"

The screenshot was of Google's newly upgraded Antigravity 2.0. The question was half-joking, but it touched on a very real transformation - the main character in programming is quietly changing.

Event Background

Antigravity is not entirely new. According to the Google Developers Blog, it first appeared on November 18, 2025, along with Gemini 3, positioned as a free, "agent-first" IDE. The so-called agent-first approach means that it doesn't focus on "you typing, AI completing" as its core, but rather the opposite: you assign tasks, and the AI agent completes them in the editor, terminal, or browser.

By the I/O conference in May 2026, Google upgraded it to 2.0, and its nature changed - from an editor to a standalone desktop application, designed around "agent scheduling and collaboration."

This Time's Focus

According to reports from the Google Developers Blog, The Next Web, and MarkTechPost, the key changes in Antigravity 2.0 are:

Independent desktop platform: no longer just an IDE, but a standalone application designed for multi-agent workflows.
New Antigravity CLI and SDK: allowing you to call and arrange agents in the command line and your own programs.
Integration of Gemini CLI: Google transitioned the original Gemini CLI into Antigravity CLI, equivalent to consolidating command-line AI capabilities into one line.
Managed Agents and enterprise support: providing managed agents in the Gemini API and enterprise-level support through the Gemini Enterprise Agent Platform.
Change in core abstraction: from "IDE assisting you in writing" to "multi-agent workflow arrangement" becoming the primary operating unit.

Market Impact Analysis

For Taiwanese users: if you're not an engineer, the significance of this event is indirect - the software and automated services you use in the future will increasingly be "created by AI agents, with humans reviewing" them. Product iterations will be faster, but it also means you need to learn to judge which ones are hastily produced by AI.

For enterprise applications: for Taiwanese software teams and startups, this is an opportunity to lower the development threshold. Originally, tasks that required several people's work, such as writing tests, modifying settings, and refactoring across files, can now be handed over to agents. However, don't misunderstand this as "being able to lay off employees": agents will make mistakes and may modify things they shouldn't when you're not paying attention, so teams actually need to shift their manpower from typing to "designing tasks and strictly reviewing" them.

For developers: this is a signal to start trying it out. Antigravity's agent-based approach, along with Cursor's background agents and various AI editors, is heading towards the same future. It's connected to the Gemini model, making it more convenient for those already using the Google ecosystem. For those who want to get into agent development, you can start by building concepts with our AI agent development guide and then try it out in practice.

Future Development Trends

By integrating Gemini CLI into Antigravity, the message is clear: Google wants to converge "command line, editor, and agent" into one suite. In the future, the competition won't be about "whose completion is more accurate," but rather "whose agent arrangement is more stable and controllable." It can be expected that in the second half of 2026, mainstream development tools will be fully agent-based, and "humans responsible for review and quality control" will become one of the core skills of engineers. The risk side shouldn't be ignored: agents running commands and modifying files autonomously require proper permission isolation and version control in formal environments, which the industry is still working on.

TheAI Academy Summary and Commentary

Antigravity 2.0 is not just another smarter completion tool, but rather a formal placement of the "AI completes, you review" model on the table.

Commentary: The value of engineers won't disappear, but it will shift - from "writing fast" to "reviewing accurately and designing well." Starting to practice "leading agents" now is more worthwhile than writing a few more lines of code.

Specific suggestions for Taiwanese developers: first, try running agents in personal projects or non-formal environments, intentionally observing where they will go wrong, and accumulating "experience from pitfalls" - this is the most valuable ability for the future. Before introducing it into formal projects, be sure to do proper version control and permission isolation.

Data Sources

(This article is compiled based on publicly available information, with functions and times following Google's official explanations.)