Google、AI開発プラットフォーム「Antigravity」を発表 Introducing Google Antigravity

Google has introduced Antigravity, a new agent-first development platform that reframes the IDE around autonomous AI agents rather than line-by-line code editing. The launch marks Google's most ambitious attempt yet to compete directly in the rapidly consolidating market for AI-native developer tools.

At the heart of Antigravity is Gemini 3, Google's latest flagship model, which drives agents that can operate across three surfaces developers typically juggle: the code editor, the terminal, and the browser. Rather than acting as a passive autocomplete or chat sidekick, these agents are designed to plan multi-step work, execute commands, and verify results in a real browser environment. Google frames the platform as a way to lift developers to a higher level of abstraction, focusing on intent and architecture while delegating mechanical implementation to agents.

Key capabilities reportedly include parallel agent execution, an Artifacts system that surfaces plans, screenshots and verifiable evidence of what an agent did, and support for long-running tasks driven by natural language. While Gemini 3 Pro is the default brain, Antigravity appears to allow selection of third-party models including Anthropic's Claude, signaling a more pluralistic stance than some competitors.

The context matters. The agentic IDE space has become fiercely contested over the past year. GitHub Copilot has rolled out an Agent Mode, Anthropic shipped Claude Code, and OpenAI continues to expand Codex and its broader agent tooling. Startups such as Cursor, Windsurf (now part of Cognition after a complex set of acquisitions involving Google talent earlier in 2024), and Replit have shown that developers are willing to abandon long-standing editors when a new tool meaningfully changes productivity. Until now, Google's developer story leaned on Gemini Code Assist as a VS Code extension, leaving a gap that Antigravity is clearly intended to fill.

There is also a strategic logic in owning the surface where coding happens. Whoever controls the IDE controls the default model, the telemetry, and the distribution channel for future agentic features. Microsoft has leveraged this through VS Code and GitHub for years, and Google appears determined not to cede that ground as software development shifts toward agent orchestration.

Several open questions remain. Reliability of long-horizon agents is still uneven across the industry, and giving an autonomous system access to a terminal and a browser raises real security and sandboxing concerns. The quality of the human-in-the-loop review experience — how easily a developer can inspect, approve, or roll back agent actions — is likely to determine whether Antigravity becomes a daily driver or a curiosity. Pricing, enterprise controls, and integration with existing CI/CD and code review workflows will also shape adoption.

If Gemini 3's reasoning gains translate cleanly into agent reliability, Antigravity could become a credible challenger to Cursor and Copilot rather than just another Google experiment. Either way, the announcement reinforces a broader industry consensus: the next phase of developer tooling will be defined less by smarter autocomplete and more by trustworthy, observable agents working alongside humans.