[1/2] Defying Gravity (AI column)
It's interesting. speech About the new IDE "Antigravity" from Kevin Howe, head of product engineering Google Antigravity at Google DeepMind. Until summer. 2025 Kevin was Head of Product Engineering at Windsurf (former Codeium)Google Acquires Windsurf Team for $2,4 Billions to create antigravity. In this talk, Howe describes the evolution of AI tools for developers as a sequence of leaps, each associated with improved models: 1Autocomplice (GitHub Copilot) 2Chat (ChatGPT with RLHF) 3Agents (Antigravity)
Antigravity launched 18 November 2025 year with Gemini 3 Pro is positioned not just as an editor with AI functions, but as an agent-oriented platform where autonomous agents become full-fledged development partners. This IDE has three components. 1Agent Manager - Central Agent Management Hub. Inbox for tasks that require attention (For example, confirmation of terminal commands)With OS notifications and the ability to manage multiple agents in parallel. 2AI Editor - VS Code fork with fast auto kit and agent sidebar. Switching between Editor and Agent Manager takes100 miss (Cmd/Ctrl+E). 3️⃣ Agent-Controlled Browser An automated Chrome that the agent uses to: - Getting context (Access Google Docs, GitHub dashboards with your authentication) - Verification of results (Clicking, scrolling, executing JavaScript to test applications) - Video recordings of actions instead of showing diff code
IDE offers a new pattern for interaction through the concept of Artifacts - these are dynamic visual representations of the agent's work: - Types of artifactsimplementation plans (PRD, product requirement definitions), task-lists, architectural diagrams Mermaid, screenshots, browser videos, walkthroughs (Final reports as PR description) **- Dynamic.**The model decides whether the artifact is needed, what type, who should see it (Other agents, conversations, knowledge base) **- Feedback system.**You can leave comments in the style of Google Docs / Figma - highlight text or areas in the images, buckle edits and send to the agent without interrupting the task.
In the past, you had to read a chain-of-thoughts sheet to track a robot, which was a bit hard. But artifacts give you a visual representation of progress -- it's like PowerPoint for presentations.
The product is built on the capabilities of Gemini 3 Pro and offers four categories of improvements 1Intelligence and reasoningbetter follow instructions, understand the nuances of using tools, cope with long tasks 2️⃣ Multimodalprocessing text, images, audio, video, code at the same time. Image caption Generation (Nano Banana Pro) integrated for iterations by design directly in the editor 3️⃣ Computer Use: Gemini variant for browser management - clicks, DOM inspection, JavaScript. The result is not diff, but video recording of actions for verification. 4Long-lived tasksAgents work in the background, notifying the need for intervention
According to Howe, the key advantage of the new IDE is symbiosis with DeepMind. The Antigravity team sits "a couple of dozen meters" away from the Computer Use team. Early access to Gemini 3 a few months before the release allowed to find the weak spots of the model and fix them in the product Antigravity used inside Google by DeepMind engineers and researchers is an example of dog fooding. This symbiosis creates a feedback loop: product shows model gaps → researchers improve model → product integrates improvements For example, Computer Use initially worked poorly until the commands synchronized data distribution and agent harness, and the artifacts concept required revisions to the model to understand the concept of a review.
In continuation We will discuss what conclusions can be drawn from the emergence of such an IDE.
#Engineering #AI #Metrics #Software #DevEx #Productivity #IDE