Google Antigravity: Hands on with our new agentic development platform

Introduction to Antigravity as an agent development platform with three key differentiators: a familiar editor with agent orchestration capabilities, browser actuation powered by Gemini 3.0, and artifacts as indivisible task units for developer-agent interaction. The platform enables agents to operate with less human intervention for longer periods.

• •Agent manager allows orchestration of multiple agents operating over codebases
• •Browser actuation enables agents to verify work and perform research tasks
• •Artifacts provide verifiable steps that developers can review and iterate on
• •Gemini 3.0 enables agents to execute longer tasks with minimal human intervention

Discussion of the progression from GitHub Copilot's autocomplete (2021) to chat experiences to agentic IDEs. Emphasis on how modern coding agents need multimodal capabilities beyond just code understanding - including browser interaction, research, and document analysis - to truly accelerate software development.

• •Industry evolution: autocomplete → chat panels → agentic experiences → multimodal agents
• •Models must understand more than code: browser, papers, bug reports, system designs
• •DeepMind's frontier model capabilities across multiple dimensions enable ambitious product features
• •Next-generation tools require understanding of non-code surfaces where developers work

Analysis of how code writing represents only ~20% of developer time. Antigravity addresses the full hierarchy of developer needs: what to build, how to build it, and building it. The platform integrates with surfaces like Google Docs for system designs and bug tracking systems to provide comprehensive development assistance.

• •Developers spend time writing, reviewing, testing, debugging, and deploying code
• •Understanding system designs and documentation is critical for large-scale changes
• •Operating only on code limits acceleration to ~20% of developer workflow
• •AI can help with 'what should I build' and 'how should I build it' questions, not just implementation

Antigravity is heavily dogfooded internally at Google DeepMind, including by leadership like Sergei. The internal Google codebase provides unique hardening for enterprise-scale complexity. Target users include researchers and developers building complex applications, with the goal of enabling singular developers to build entire companies.

• •Google's massive codebase hardens the product for largest enterprise use cases
• •Internal researchers provide feedback on solving very complex tasks beyond simple apps
• •Researchers need agents to handle training jobs, evals, and analysis - not just code
• •Vision: enable individual developers to build their dreams and entire companies

Discussion of the complementary relationship between AI Studio (for vibe coding and quick prototypes) and Antigravity (for complex, production applications). A migration path will enable users to start in AI Studio and move to Antigravity as projects mature, making apps more economically valuable.

• •AI Studio ideal for quick vibe coding with AI features built-in
• •Antigravity designed for iterating on applications over time and adding features
• •Migration path from AI Studio to Antigravity will be available
• •Transition enables technical users to make deeper architectural decisions for scale

Exploration of how the traditional editor won't disappear but will evolve. Developers will spend less time on individual lines of code and more time reviewing artifacts - higher-level abstractions of agent work. The editor becomes one of many work surfaces rather than the primary interface.

• •Editor remains necessary for developers to see and control underlying code
• •Time spent writing individual lines will decrease; artifact review will increase
• •Editor becomes one work surface among many (code review, docs, sheets)
• •Artifacts serve as the interface between developers and agents for code review

Antigravity provides configurable autonomy levels through 'agent assisted development' mode. Agents decide when to notify users based on task importance, while developers can provide asynchronous feedback like comments on Google Docs. The system balances long-running tasks with iterative human feedback.

• •Configurable autonomy levels let agents decide when to notify users
• •Asynchronous feedback mechanism allows comments anywhere, anytime
• •Agents fold in feedback at appropriate times during execution
• •Users can review work after extended periods (coffee breaks) and provide corrections

Discussion of how Antigravity currently runs locally but will evolve to support server-side execution for tasks outliving developer machines. Emphasis on the importance of iterative feedback even for asynchronous tasks, as developers learn requirements through seeing code rather than specifying everything upfront.

• •Current local execution will converge with server-side background agents
• •Models capable of long execution can also handle short tasks effectively
• •Iterative feedback remains critical even for multi-day tasks
• •Developers discover requirements through seeing code, not just upfront specification

Antigravity includes a knowledge panel that builds up understanding of user preferences and context over time, similar to Google Search's contextual understanding. The system learns from conversations to avoid re-deriving information and captures knowledge that exists outside the codebase.

• •Knowledge panel learns from usage patterns to provide contextual assistance
• •System remembers preferences without re-deriving on each interaction
• •Captures knowledge from meetings and discussions not in codebase
• •Model decides when to save meaningful state for future retrieval

While Antigravity's principles could generalize to many work domains, the team maintains focus on developers to avoid building a product that's great for no one. The number of developers is growing as building becomes easier, from Assembly to Python to AI-assisted development.

• •Core capabilities (browser, research, artifacts) applicable beyond coding
• •Team uses Antigravity for non-coding tasks like analyzing sports statistics
• •Deliberate focus on developers to maintain product excellence
• •Developer population growing as barriers to entry decrease with each abstraction layer

Antigravity integrates Google's full suite of state-of-the-art models including Gemini 3.0 for coding, Imagen for image generation, and UI control capabilities. The platform provides cutting-edge capabilities on day one and enables the team to push both product and research frontiers.

• •State-of-the-art models across image generation, UI control, and coding
• •Latest Gemini capabilities embedded into product from day one
• •Team positioned to push both product features and model research
• •Ensemble of models creates powerful multimodal development experience

Comprehensive demonstration of building an Airbnb-for-dogs app showing artifacts (task plans, images, implementation plans), asynchronous feedback, browser verification with screenshots, and parallel agent orchestration across multiple workspaces. Showcases the inbox for managing multiple concurrent agent workflows.

• •Artifacts include task descriptions, generated images, implementation plans, and verification walkthroughs
• •Asynchronous Google Docs-style comments allow feedback without blocking agent progress
• •Browser automation captures screenshots and videos as proof of verification work
• •Inbox enables orchestration of multiple agents across different workspaces simultaneously
• •Shadow editor (Command+E) provides instant access to code when needed
• •Low cost of asking questions enables firing off many parallel tasks

Recommendations for breaking work into smaller, disjoint pieces for better digestibility rather than single large tasks. Future vision includes faster experiences, more capable models handling complex tasks, and new product form factors emerging from increased model capabilities - following the pattern of autocomplete → chat → agents → agent orchestration.

• •Break features into smaller pieces for more digestible implementation plans and verification
• •Ensure parallel tasks are disjoint to avoid coordination issues
• •Model improvements will enable new product form factors beyond current capabilities
• •Expect significantly faster experiences and ability to handle more complex tasks
• •Pattern: as models improve, new human-AI interaction paradigms emerge