Defying Gravity - Kevin Hou, Google DeepMind

Key Concepts

• Anti-gravity: A new AI developer platform from Google DeepMind, designed with an "agent-first" philosophy.
• Agent-First Philosophy: Prioritizing AI agents as the primary interaction model for software development.
• Agent Manager: The central hub of Anti-gravity, providing an agent-centric view and managing multiple agents.
• AI Editor: A feature-rich code editor with lightning-fast autocomplete and an agent sidebar, mirroring the Agent Manager.
• Agent-Controlled Browser: An integrated Chrome browser that allows agents to interact with the web, retrieve context, and perform actions like clicking, scrolling, and executing JavaScript.
• Artifacts: Dynamic representations of information generated by agents, used for organization, self-reflection, communication, and inter-agent memory.
• Research-Product Flywheel: The iterative process of using Anti-gravity internally to identify model gaps and improve both the product and the underlying AI models.
• Multimodality: The ability of AI models (like Gemini 3) to process and understand various types of data, including images and text.
• Computer Vision (CV) Use: Agents' ability to interact with and understand visual elements, such as screenshots and web page renderings.

Anti-gravity: An Agent-First AI Developer Platform

This presentation introduces Anti-gravity, a novel AI developer platform developed by Google DeepMind. The platform is characterized by its "agent-first" approach, aiming to fundamentally shift how software developers interact with AI tools. Kevin How, leading the product engineering team at Google Anti-gravity, details the platform's architecture, core functionalities, and the underlying principles that drove its development.

Core Components of Anti-gravity

Anti-gravity is built around three primary surfaces:

1. Agent Manager:

   • Functionality: Serves as the central control panel and "agent-first view." It elevates the user's perspective beyond just code diffs to a higher-level overview of agent activities.
   • Key Features:
     • Single Window: Only one Agent Manager window is active at a time.
     • Inbox: Manages tasks requiring user attention, such as running specific terminal commands, which are surfaced for explicit approval.
     • OS-Level Notifications: Alerts users to critical agent activities.
     • Parallelism and Orchestration: Designed to handle multiple projects and concurrent agent tasks.
   • Purpose: To optimize the UI for "artifacts" and manage multiple agents efficiently.

2. AI Editor:

   • Functionality: A comprehensive code editor offering standard features like lightning-fast autocomplete.
   • Key Features:
     • Agent Sidebar: Mirrors the Agent Manager, allowing for focused work within the editor.
     • Instant Hops: Seamless and rapid (under 100 milliseconds) switching between the editor and the Agent Manager using keyboard shortcuts (Command E or Control E).
   • Purpose: To provide a familiar and powerful environment for developers to complete 80-100% of their tasks, with seamless integration with agent functionalities.

3. Agent-Controlled Browser:

   • Functionality: An integrated Chrome browser that grants agents access to the richness of the web.
   • Key Features:
     • Context Retrieval: Leverages existing browser authentication (e.g., Google Docs, GitHub dashboards) for agents to access relevant information.
     • Agent Control: Agents can actively control the browser, including clicking, scrolling, and executing JavaScript, enabling testing and interaction with web applications.
     • Verifiable Results: Demonstrates agent actions through screen recordings, with a visual indicator (blue circle) showing mouse movement. This provides tangible proof of task completion.
   • Example: A demo showcased an agent generating random artwork by refreshing a webpage, with the browser interaction recorded for verification.

The "Agent-First" Paradigm Shift

The core innovation of Anti-gravity lies in its paradigm shift: agents now operate outside the traditional IDE and can interact across multiple surfaces. This allows for a more dynamic and integrated development workflow. The platform is designed to leverage advancements in AI model capabilities, particularly with the recent release of Gemini 3 Pro.

Driving Principles and Industry Trends

The development of Anti-gravity is rooted in observing industry trends and product principles:

• Model Capability Evolution: The platform's design is motivated by step-function improvements in AI models, moving from autocomplete (Copilot) to chat, and now to agents.
• DeepMind Integration: Being embedded within DeepMind provides early access to cutting-edge models like Gemini, allowing the team to identify strengths, exploit capabilities, and address model gaps.
• Raising the Ceiling of Capability: Anti-gravity aims to push the boundaries of what AI agents can achieve, particularly in areas like browser interaction and multimodal understanding.

Key Improvements Powered by Gemini 3

Anti-gravity capitalizes on four main categories of improvements driven by advanced models like Gemini 3:

1. Intelligence and Reasoning:

   • Enhanced Instruction Following: Models are better at understanding and executing complex instructions.
   • Nuanced Tool Use: Improved ability to utilize various tools, including the browser and JavaScript execution.
   • Longer-Running Tasks: Agents can now handle more extended processes, allowing for background execution and deeper thinking.

2. Multimodality:

   • Image Understanding: Gemini 3's advanced multimodal capabilities are crucial for verifying screenshots, recordings, and interacting with visual elements.
   • Synergistic Nature: The platform leverages both Gemini 3 Pro and image generation models for a richer developer experience.

3. Browser Use as a Major Unlock:

   • Beyond Code: Browser interaction addresses the "how to build it" aspect of development, providing context from bug dashboards, experiments, and institutional knowledge.
   • Verification: Screen recordings of agent actions in the browser serve as a powerful verification mechanism.
   • Example: A flight tracker demo illustrated an agent using the browser to input flight IDs and retrieve information, with the entire process recorded.

4. Image Generation and Mockups:

   • Design Iteration: The platform supports iterating on designs directly in image space, allowing for comments and updates similar to collaborative tools.
   • Figma-Style Interaction: Enables highlighting and commenting on visual mockups, with agents intelligently incorporating feedback.
   • Example: A demo showed an agent taking design comments on a mockup and updating it accordingly.

Artifacts: A New Interaction Pattern

A significant innovation in Anti-gravity is the introduction of artifacts.

• Definition: Artifacts are dynamic representations of information generated by an agent. They are crucial for organization, self-reflection, communication, and memory.
• Purpose:
  • Organization: Keep agents structured and manage their outputs.
  • Self-Reflection: Aid agents in understanding their own processes.
  • Communication: Provide users with visual outputs like screenshots or screen recordings.
  • Inter-Agent Memory: Facilitate information sharing between different agents or conversations.
• Dynamic Nature: Agents decide if an artifact is needed and what type of artifact to generate, making the process adaptive.
• Common Artifact Types:
  • Markdown: Used for plans, walkthroughs, and task lists.
  • Plans: Similar to Product Requirements Documents (PRDs), outlining steps and open questions.
  • Walkthroughs: Demonstrations of how an agent completed a task, akin to a PR description.
  • Images: Visual outputs.
  • Screen Recordings: Visual logs of agent actions.
  • Mermaid Diagrams: For visualizing structures and processes.
• User Interaction with Artifacts:
  • Feedback Loop: Users can provide feedback on artifacts through comments, similar to Google Docs or GitHub.
  • Iterative Refinement: Agents incorporate user comments to refine their work without interrupting the execution loop.
  • Example: Comments on a design mockup or highlighted text in a markdown plan are used to guide the agent's next steps.

The Research-Product Flywheel: Building for Ourselves

A core philosophy behind Anti-gravity's development is the research-product flywheel.

• Internal Dogfooding: Google engineers and DeepMind researchers are the primary users of Anti-gravity. This internal usage provides invaluable feedback for identifying model gaps and product improvements.
• Synergy: This close integration between product development and research teams allows for rapid iteration and pushes the frontier of AI capabilities.
• Examples:
  • Computer Vision Use: Collaboration with the CV team identified mismatches in data distribution and agent harness issues, leading to improvements on both sides.
  • Artifacts Development: Initial artifact capabilities were refined through close work with the research team, leading to Gemini 3 Pro's enhanced ability to handle them.
• Outcome: Anti-gravity is positioned to be the most advanced product on the market because it is built by and for its own users, creating a self-reinforcing cycle of innovation.

Conclusion and Future Vision

Anti-gravity represents a significant step forward in AI-powered software development. Its agent-first approach, integrated surfaces, and the innovative artifact system aim to streamline workflows and enhance developer productivity. The platform's development is driven by a commitment to pushing AI capabilities through a unique research-product flywheel. The team is excited to bring Anti-gravity to market and welcomes user feedback to continue its evolution. The vision is a future where developers spend more time in the Agent Manager, orchestrating parallel sub-agents and leveraging the full power of AI.