Hands on with our new agentic development platform

Key Concepts

• Anti-gravity: An agent development platform with an IDE, agent manager, and browser actuation capabilities.
• Agent Manager: Allows orchestration of multiple agents operating over a codebase.
• Browser Actuation: Agents can interact with and control the browser, enhanced by Gemini 3.
• Artifacts: Indivisible units of tasks and interactions between agents and developers, providing verifiable steps.
• Agent Assisted Development: Feature allowing developers to specify the level of autonomy for agents.
• Knowledge Panel: A feature where the product builds knowledge of user usage over time to improve future interactions.
• Generative Media Models: A suite of Google's models, including image generation and latest generative AI models, integrated into Anti-gravity.
• Vibe Coding: A rapid, intuitive approach to building applications, often starting with AI Studio.
• Dogfooding: Internal use of a product by its developers to gather feedback and iterate.

Anti-gravity: An Agent Development Platform

Product Overview and Key Differentiators

Anti-gravity is presented as an agent development platform designed to enhance developer productivity. Its core components include a familiar IDE, an agent manager, and a novel browser interface. A key differentiator is the agent manager, which enables the orchestration of numerous agents capable of operating directly on a developer's codebase. Furthermore, with the advent of Gemini 3, agents can now actuate the browser in a highly capable manner.

A central concept in Anti-gravity is "artifacts." These are described as indivisible units of tasks and interactions, serving as verifiable steps that agents undertake. Developers can engage with these artifacts iteratively, allowing for the completion of increasingly complex tasks. The inspiration behind this approach stems from the observation that advanced models, like Gemini 3, can now handle longer sequences of tasks with reduced human intervention, a capability Anti-gravity aims to maximize for developers.

Evolution of Developer Tools and Agentic Experiences

The discussion traces the evolution of developer tools, starting with GitHub Copilot in 2021, which offered basic autocomplete. This was followed by chat-based experiences (like Gemini and ChatGPT) integrated into IDEs, providing personalized code understanding. The current frontier, according to Verun Mohan, is "agentic IDEs and agentic experiences." This next phase necessitates models that can go beyond just understanding code, extending to actions like browsing the web, conducting research, and analyzing bug reports to identify tasks. This multi-dimensional capability of models is a key driver for Anti-gravity's development.

Defining the Developer's Role and Scope of AI Assistance

Writing code is acknowledged as only a small part of a developer's overall responsibilities. The broader scope includes code review, testing, debugging, and deployment. Developers often utilize various surfaces for these tasks, such as Google Docs for system designs. Anti-gravity aims to address this by operating beyond the confines of just code, acknowledging that understanding system designs is crucial for large-scale code changes or reviews.

The analogy of "chopping wood" is used to illustrate the developer's time allocation. If coding represents 20% of a developer's time, accelerating only that portion has limitations. Anti-gravity seeks to accelerate the "how to build it" aspect, which involves more complex decision-making and research, areas where AI can provide significant acceleration.

Target User Persona and Internal Adoption (Dogfooding)

A strong emphasis is placed on "dogfooding" the product internally at Google. This practice provides rich feedback for iteration before public release. Google's internal codebase and infrastructure are noted as being exceptionally complex, presenting a unique challenge and hardening the product for large enterprises.

While the primary target user is the developer, aiming to empower them to build complex applications and even entire companies, it's observed that non-technical individuals are already using Anti-gravity internally. The product strategy aims to "raise the ceiling" for advanced users while simultaneously making it accessible to those who are "technically adjacent." The concern is to avoid becoming a product that is "good for neither persona."

Integration with AI Studio and Future Paths

There's a planned integration path from AI Studio to Anti-gravity. AI Studio is positioned as an excellent platform for "vibe coding" and rapidly building initial applications. For developers who wish to iterate and add more features over time, Anti-gravity is intended to be the right platform. This transition is expected to be a "magical experience," making applications more economically valuable.

The Future of IDEs and Agent Orchestration

The discussion posits that the traditional IDE is unlikely to disappear in the foreseeable future, as developers require control to see and edit code at the lowest level. However, the time spent writing individual lines of code and relying on autocompletes is expected to decrease significantly. The editor may evolve into another work surface alongside code review tools and document editors, all of which agents can operate on and collaborate with humans.

The focus is shifting towards "artifacts" and verifiable units from agents, with the expectation that developers will spend more time reviewing these. As AI takes on more code generation, the challenge becomes enabling developers to review this code efficiently without writing every piece themselves. This necessitates a higher-level abstraction between the developer and the agent, where the editor or lines of code might not be the primary interface.

Human-in-the-Loop, Verification, and Autonomy

Anti-gravity incorporates "agent assisted development," allowing developers to specify the level of autonomy granted to agents. Agents can decide whether to notify the user about terminal commands or tasks based on a defined threshold. Trust in the AI's judgment is balanced with an "escape hatch" for developer feedback.

Feedback can be provided asynchronously, similar to comments in a Google Doc, directly on artifacts or in the chat panel. The agent will integrate this feedback when appropriate. This allows for a mix of autonomous operation and human intervention, enabling developers to review work done while they were away and provide iterative feedback.

Local vs. Server-Side Execution and Asynchronous Experiences

The current implementation of Anti-gravity runs locally on the developer's device, connected to server-hosted models. This requires the developer's machine to be active for progress. The convergence of local and server-side execution is anticipated, especially as models become capable of handling longer-duration tasks.

The importance of asynchronous experiences is highlighted, but the challenge lies in effective communication of intentions to the model. Developers often have implicit requirements not explicitly stated. This iterative journey, where insights are gained during the building process, necessitates mechanisms for users to understand progress and iterate.

Contextual Understanding and Self-Improvement (Knowledge Panel)

A key feature is "self-improvement," built as a primitive into Anti-gravity. The product builds knowledge of user usage over time, reducing the need to re-derive information. This "knowledge panel" captures intrinsic knowledge from interactions, including those outside the codebase (e.g., meetings), which can bridge the gap when the codebase might be slightly out of date regarding user intent.

Despite these advancements, an interactive experience remains crucial. The analogy of an engineering team delivering a feature after weeks of work, only for it to be incorrect, underscores the need for continuous check-ins and iterative processes, mirroring human best practices.

Generalization Beyond Coding and the "Builder" Persona

The capabilities demonstrated in Anti-gravity, such as human-in-the-loop components, browser interaction, and research, suggest a broader generalization beyond just software development. The underlying intelligence of the models enables them to solve a wide range of tasks.

While Anti-gravity is currently focused on developers, the potential for broader application is evident. The strategy is to avoid diluting the product by trying to be "great for everyone," instead focusing on deeply understanding and serving the developer community. The growing number of "builders and developers" worldwide, driven by increasing ease of building, is a key trend. Anti-gravity aims to lower the barrier to entry for new builders while simultaneously increasing the ceiling for experienced developers, particularly in complex codebases.

Naming and Acronyms

The name "Anti-gravity" was chosen externally, reflecting a desire to create an experience where developers are not limited, with imagination being the only constraint. The space theme and the idea of "no limits" were attractive. The acronym "AGY" is noted as being short, catchy, and a subtle nod to AGI (Artificial General Intelligence).

Model Suite and Capabilities

Anti-gravity leverages a suite of Google's state-of-the-art models, not just a single model like Gemini 3 Pro. This includes generative media models, image generation capabilities, and future advancements. The integration of these diverse models allows Anti-gravity to access and utilize capabilities across different product experiences where they can be maximally helpful. This ensemble approach is seen as powerful, pushing model capabilities in areas like UI control and complex agentic workloads.

Demo: Airbnb for Dogs Application

A demonstration of Anti-gravity involved building a simple "Airbnb for dogs" application. The process showcased several key features:

• Design Generation: The agent generated design concepts (images) for the application, such as "K9 Concierge" and "Pup Stay."
• Artifacts and Iterative Feedback: Images were presented as artifacts. The developer provided feedback, requesting a name change from "K9 Concierge" to "Pup Stay," which the agent incorporated. This feedback was asynchronously provided in a Google Doc-like format.
• Implementation Plan and Documentation: The agent generated an implementation plan and created documented HTML structure, incorporating the user's feedback regarding documentation style.
• Knowledge Panel Integration: The concept of the knowledge panel was discussed, where consistent feedback (e.g., requesting documentation or specific styles) would be learned and potentially used automatically in the future.
• Installation and Local Changes: When the agent needed to install something that would modify the local machine, it prompted the user for acceptance, allowing them to review the code changes in the editor.
• Browser Actuation: The agent demonstrated the ability to actuate the Chrome browser, creating screenshots of the running application and providing a walkthrough of its actions. This included highlighting the agent's operational area on the page.
• Parallel Agent Execution: The demo showcased the ability to create multiple agents operating in parallel on different tasks within the same or new workspaces. For example, filling in the "Experiences" and "Become a Host" tabs for the Airbnb app, and independently analyzing a "Baseball Stats" project.
• Inbox and Workspace Management: The "inbox" serves as a central hub for all agent workstreams, regardless of the workspace. Workspaces can be decoupled from the editor, allowing for a decluttered experience.
• Verification and Proof of Work: The agent provided a walkthrough with screenshots and recorded actions, demonstrating the verification of its work, such as implementing the "Become a Host" page.
• Model-Driven Iteration: The developer found it enjoyable to break down tasks into smaller pieces and let the agent work on them independently, finding the implementation plans and verification more digestible.

Future Outlook and Product Evolution

The current version of Anti-gravity is considered V0.1, with significant future development planned. The pace of progress is expected to be rapid, driven by increasing model capabilities. New product form factors, such as artifacts, will emerge as models become more capable of handling longer and more complex tasks. The agent manager, enabling the orchestration of numerous agents, is a direct result of models being able to execute tasks for extended periods. The future promises faster experiences, more capable and complex tasks performed by models, and increased opportunities for developers to step away while their software is being built.