TLDR Anti-gravity is an agent-first IDE that lets you run multiple AI agents in parallel via an editor, agent manager, and a browser agent, all while keeping data private and local. It uses artifacts and implementation plans to capture context, supports Git integration and model switching (Gemini 3 Pro, Claude Sonnet 4.5), and includes an inbox for human-in-the-loop decisions. The demo shows building a local RSS reader in Next.js/TypeScript with a sandboxed browser, parallel tasks, and local LLMs, highlighting privacy-first operation and a cohesive flow between editor, manager, inbox, and browser.
Install and harden anti-gravity for privacy-first setup
Start by installing anti-gravity and configuring privacy and telemetry settings. Create a sandboxed browser profile to isolate data during AI-driven testing. Use local-first operation to keep sensitive information on your machine and avoid unnecessary data exposure. This foundation makes subsequent workflows safer, repeatable, and easier to audit. A solid setup reduces friction when you scale agent workflows later.
Ground AI reasoning with artifacts and implementation plans
Rely on artifacts and implementation plans to capture context, outline how to build, and guide AI decisions. Use targeted comments on specific UI elements to steer the agent’s work without silencing its autonomy. These mechanisms help maintain traceability and reduce misinterpretations as tasks scale. By tying actions to concrete plans, you make outcomes more predictable and auditable. This approach supports better collaboration with human reviewers.
Plan first, then execute: switch between planning and fast modes
Use planning mode to map out the approach before committing to implementation. When confident, switch to fast mode to accelerate execution while keeping an eye on results. Choose models like Gemini 3 Pro or Claude Sonnet 4.5 Thinking to balance capability and cost. The ability to toggle between modes helps you adapt to evolving requirements. This stepwise rhythm keeps complex projects orderly and auditable.
Coordinate parallel work with the agent manager
Under the agent manager you can run multiple agents in parallel, with different roles handling planning, state management, or data gathering. The inbox shows actions that require human input, so you stay in the loop without micromanaging. Live agent steps are visible in the editor, terminal, and browser, helping you track progress and quality. This orchestration reduces bottlenecks and speeds up delivery. Be mindful of potential file conflicts and coordinate task boundaries.
Harness the browser agent in a sandbox to test and verify
The browser agent controls a sandbox Chrome to navigate websites, click elements, and pull data from feeds like Hacker News, Verge, and the New York Times. It can generate proof of work via screen recordings, providing tangible validation of automated tasks. This setup supports local-first privacy while enabling real-world testing. Use the browser agent to prototype UI flows and confirm behavior before committing changes. Privacy-conscious design helps you avoid leaking data when testing.
Integrate with Git and manage workspaces for versioned progress
Spawn new workspaces and connect Git to track changes across experiments. Automatic Git commit messages are generated to document what the agent built or changed, improving traceability. The system supports workspace-level isolation so multiple experiments don’t collide. Regular commits and clear branch management make collaboration smoother. This practice aligns AI-assisted development with standard software workflows.
Keep humans in the loop and plan for resilience
The inbox surfaces blocked actions awaiting human approval, enabling a safe human-in-the-loop workflow. A daily briefing agent can summarize state and upcoming steps, helping teams stay aligned. If a model quota runs out or an issue arises, you can switch models and continue with the flow. The design emphasizes transparency of agent actions and a clear audit trail. This resilience is particularly valuable in first-release environments.
What is anti-gravity?
Anti-gravity is an agent-first IDE designed around agentic workflows. It includes an editor, an agent manager, a browser agent, and an inbox to run multiple agents in parallel, while keeping you in the loop. It uses artifacts and implementation plans to capture context, plan how to build, and guide AI with targeted comments. It supports planning and fast modes, model options like Gemini 3 Pro or Claude Sonnet 4.5 Thinking, and can spawn new workspaces with Git integration.
What are the main components of anti-gravity?
The main components are an editor, an agent manager, a browser agent that can test and fix your app, and an inbox. It enables running multiple agents in parallel, switching between planning and fast modes, and using implementation plans and artifacts to guide development, with Git integration and the ability to spawn new workspaces.
How does the parallel agent capability work?
You can run multiple agents in parallel under the agent manager. Different agents can handle distinct tasks concurrently (for example, one managing state and a daily briefing while another handles local LLM integration), with safeguards to avoid file conflicts.
What models are supported and how do you switch if quotas run out?
You can choose models such as Gemini 3 Pro or Claude Sonnet 4.5 Thinking. If a model quotas out or you encounter debugging issues, you can switch to another model and continue, with features like follow mode showing live agent steps in the editor, terminal, and browser.
How does the browser agent work?
The browser agent controls a separate sandbox Chrome instance to navigate websites, click UI elements, pull data from feeds (like Hacker News, Verge, and the New York Times), and generate proof-of-work recordings.
What happened in the hands-on walkthrough?
Callum installed anti-gravity, configured privacy and telemetry, and set up a sandboxed browser profile to isolate data. He built a local RSS reader with Next.js, TypeScript, and Zustand, used an implementation plan and inbox for human-in-the-loop decisions, and moved the project into a workspace with Git versioning.
How does human-in-the-loop work in anti-gravity?
An implementation plan and an inbox provide human-in-the-loop decisions. Actions can be blocked and require human approval before proceeding.
How is versioning and Git integration handled?
Projects can be moved into a workspace with Git versioning, and Git integration is supported to manage changes.
How does anti-gravity address privacy and local-first operation?
The platform emphasizes privacy and local-first operation, arguing that anti-gravity can run locally and keep data private, unlike Google AI Studio which is sandboxed and lacks persistent local back-end capabilities.
What automation features are included?
Automation includes automatic Git commit message generation, walk-through artifacts that narrate what was built, and an inbox that shows blocked actions awaiting human approval.
What about UI improvements and Nanobanana Pro?
There is potential for UI mockups via Nanobanana Pro integrated into anti-gravity.
Who created anti-gravity and how can you provide feedback?
The creator is Callum, aka Waterloots. He invites feedback in the comments and positions anti-gravity as part of a broader vibe-coding ecosystem, comparing it favorably to prior tools while acknowledging some bugs typical of a first release.
"Coffee and a Mike" with Martin Armstrong | SOLUTIO...
What Is Google Antigravity? 🚀 AI Coding Tutorial & ...
The most overlooked supply and demand rule in stock...
How to Get UNLIMITED Leads For Cold Email in 2026...
How I Used ABM to Scale to $13M ARR in 36 Months...
GPT 4.5 - not so much wow...