- Identity & Core Mission - Who it is and how it operates.
- Tools Reference - A guide to its capabilities and toolset.
- Workflow & Agentic Mode - Understanding the "Task View" and artifacts.
- Browser Automation - Details on the browser subagent.
- Reverse engineer Antigravity and learn things in reverse order.
- Explore agent communication and implementation details.
- Explore browser automation.
Antigravity is an agentic AI coding assistant designed by the Google Deepmind team and released on 18th November 2025.
- Workspace: It can only access files in the specified active workspaces (e.g.,
/home/ravi/Desktop/antigravity). - Internal Memory: It uses
/home/ravi/.geminifor internal state and artifacts, but avoids writing project code there. - Communication: When in "Agentic Mode" (focusing on a task), it communicates primarily through
notify_userto minimize noise, only surfacing for reviews or critical questions.
Antigravity possesses a suite of tools to interact with the file system, terminal, and browser.
view_file/view_file_outline: Read file contents and structure.write_to_file: Create new files.replace_file_content/multi_replace_file_content: Precise editing of existing files.list_dir: Explore directory contents.find_by_name/grep_search: Innovative search capabilities.
run_command: Execute shell commands (requires user approval).command_status/read_terminal: Monitor command output.send_command_input: Interact with running processes.
codebase_search: Semantic search for code snippets.notify_user: The primary channel for updates and review requests during active tasks.task_boundary: Manages the "task view" UI to group related actions.
read_url_content: Fetch static web content.browser_subagent: Full browser control for complex interactions.
Antigravity operates in a structured "Agentic Mode" to handle complex tasks without overwhelming the user.
- Task Boundary: Defines a high-level task using
task_boundary. This creates a UI block in the chat. - Iterative Work: Performs file edits, command runs, and analysis.
- Updates: Periodically updates the task status to reflect progress.
- Completion: Uses
notify_userto exit the task mode and return control to the user.
It maintains persistent documents to track state across the conversation. These live in a dedicated "brain" directory (/home/ravi/.gemini/antigravity/brain/...).
task.md: A living checklist of the current mission.implementation_plan.md: Design documents created during the PLANNING phase.walkthrough.md: Proof-of-work documents created during the VERIFICATION phase to show what was tested.
- PLANNING: Researching and designing. Output:
implementation_plan.md. - EXECUTION: Writing code. Output: Source code changes.
- VERIFICATION: Testing changes. Output:
walkthrough.md.
For tasks involving web interaction, Antigravity can deploy a specialized Browser Subagent.
- Navigation: Open URLs, click links, navigate usage.
- Interaction: Click buttons, type text, fill forms.
- Observation: Read the DOM, capture screenshots.
- Recording: All sessions are recorded as videos (saved as artifacts).
It uses the browser_subagent tool. This agent has its own limited set of tools specifically for browser control.
- The subagent is a separate entity; Antigravity provides it with a high-level task.
- It is best for verification (e.g., "Login and check if the dashboard loads") or data extraction.