Google Antigravity, 2025: the editor that decides for itself
In a demo from the week Google Antigravity launched, a tester asked the editor to build a flight tracker. The agent drafted a plan, wrote the HTML and JavaScript, ran a local server in the terminal, opened the built-in browser, took a screenshot of what it had rendered, spotted a CSS misalignment in its own screenshot, corrected it, and committed. The tester had typed one sentence. No one specified steps two through nine.
Google launched Antigravity on November 18, 2025, four months after a hiring deal that brought Varun Mohan — CEO of Windsurf — and co-founder Douglas Chen into Google DeepMind. Google had licensed Windsurf’s technology for $2.4 billion that July, after an OpenAI acquisition at $3 billion collapsed over IP routing rights. By November, the new team had shipped their first product under Google’s banner: a development environment that looked broadly like an editor but worked like something else.
The interface split into two surfaces. The Editor view was familiar — code completion, file tree, integrated terminal. The Manager view was not: a canvas for spawning multiple agents, each working in its own workspace, running asynchronously while the developer did something else entirely. Antigravity generated Artifacts alongside its code — task lists, implementation plans, browser recordings — so that every autonomous action left a paper trail. Google called the design “verifiable”; an agent that documented its own reasoning was easier to trust than one that simply made changes and waited for you to notice.
The editor gave you two ways to configure that trust. Turbo mode let the agent execute terminal commands on its own authority. Strict mode sandboxed every command and required explicit sign-off before anything ran. In practice, most users started in Strict and switched to Turbo once they understood what the agent was likely to do — which is to say, once they had lost their certainty that they did.
Within three weeks of launch, the trust question grew a sharper edge. A developer asked the agent to clear its cache. The agent misread the scope of the command and erased the contents of an entire drive. Aaron Portnoy, a security researcher, disclosed that a compromised dependency could direct the agent to install persistent backdoors — because the agent, by design, could run arbitrary terminal commands. In April 2026, Pillar Security found code execution vulnerabilities that persisted even in Strict mode. Google patched them.
None of this stopped the adoption. The platform was free at launch, and early users settled on a description that stuck: the tool was like “hiring a talented but inexperienced junior developer: incredibly fast, occasionally reckless, and in need of supervision.” Every code editor before it — vi, Emacs, VS Code, Cursor — had been a surface for a human developer’s decisions. Antigravity was a surface for the agent’s decisions, with the developer watching.
Every editor since TECO had answered the same underlying question: how do you give a human precise control over text? Antigravity answered a different question. It assumed that problem was already solved — or that you no longer cared about solving it yourself — and asked instead: how much of the decision-making do you want to hand over, and under what terms?
The interface for that answer is still being negotiated.
Sources
- Google Antigravity — Wikipedia — launch date, Windsurf licensing deal, product architecture, security incidents, Antigravity 2.0.
- Google hires Windsurf CEO Varun Mohan in $2.4 billion deal — CNBC — structure of the acquihire, Mohan and Chen’s roles at Google DeepMind.
- Exploring Google Antigravity: first impressions — PromptLayer Blog — Manager view, Artifacts system, flight tracker demo, “junior developer” characterisation, early reliability issues.