This video introduces Antigravity, a new agentic development platform from Google DeepMind. Varun Mohan explains that the platform integrates a familiar IDE with an "agent manager" for orchestrating multiple agents and a browser actuation capability powered by Gemini 3.0. A key concept discussed is "artifacts," which are verifiable work units (like design mockups, implementation plans, or test results) that facilitate communication and asynchronous feedback between the developer and the AI agents. The platform is designed to handle the entire developer workflow, which extends beyond writing code to include research, design, and testing. The discussion covers the balance between agent autonomy and human oversight, the use of a "knowledge panel" for personalization and context retention, and the future evolution of the product as AI models become more capable. A demo showcases an agent building a simple web application, responding to feedback, and verifying its work in a live browser.
Antigravity: Google DeepMind's New Platform for Agentic Software Development
0:22 An Agent Development Platform: Antigravity is not just an IDE. It combines a familiar editor with an agent manager to orchestrate multiple AI agents and a browser surface for agents to perform tasks like research and UI testing.
3:51 Beyond Just Coding: The platform aims to accelerate the entire software development lifecycle, not just code generation. This includes tasks developers regularly perform outside the editor, such as reviewing design documents, researching solutions, and debugging.
13:56 Human-Agent Collaboration: Antigravity allows developers to set the level of autonomy for agents. The system is designed for an iterative, asynchronous workflow where developers can provide feedback on "artifacts" at any time, which the agent will incorporate into its process.
19:00 Self-Improving Context with the Knowledge Panel: As a developer uses the platform, the agent learns their preferences, coding style, and project-specific knowledge. This information is saved to a "knowledge panel," allowing the agent to provide more relevant and accurate assistance over time without needing to be re-instructed.
28:46 Artifacts as a Communication Tool: Artifacts are the primary way agents communicate their work to the developer. These are self-contained, verifiable outputs like design mockups, implementation plans, generated code, or videos of browser tests. Developers can comment on these artifacts to guide the agent.
34:22 Live Browser Actuation: A core feature is the agent's ability to control a Chrome browser instance. This allows it to perform UI testing, verify that a web application is working correctly, and even conduct research, providing screenshots and recordings as proof of its work.
38:03 Orchestrating Multiple Agents: The "agent manager" allows a developer to run multiple agents in parallel, even across different projects. This enables multitasking, such as having one agent build a new feature while another analyzes a separate codebase.
43:31 The Future is More Capable Agents: The platform is considered V0.1. As underlying models like Gemini improve, the product will evolve to support longer, more complex tasks, making the development experience faster and more automated.