Google Gemini Complete Guide

Antigravity Architecture Deep Dive - Design Philosophy, Governance UI, and Security Settings

Target Audience: Engineers evaluating migration to AI IDEs, and those interested in governance for agent-driven development.

Key Points

• Specific Governance UI Settings Covers security setting names and design principles confirmed in official documentation, including Terminal Sandboxing, Browser URL Allowlist, and Strict Mode.
• Manager View Operations Explains the UI structure of Agent Manager (Mission Control), which exists as a peer to Editor View, based on official codelab information.
• Browser Subagent Authentication and DOM Interaction Documents the separate Chrome profile design, Gemini 2.5 Pro UI Checkpoint-based operation model, and video recording mechanisms.

As of March 2026, AI-driven development tools are each evolving through different approaches. Cursor, Claude Code, GitHub Copilot — each extends agent capabilities in its own way. Google Antigravity takes a design centered on integrating management, evidence, verification, and review into a cohesive unit.

Individual features can be achieved with other tools as well. What characterizes Antigravity is that these are built into the product as a single operational unit from the start¹. This guide organizes the design philosophy based on official documentation and codelabs, covering setup through security configuration.

Architecture and Design Philosophy

Let's start by looking at Antigravity's UI structure.

graph LR
    subgraph "Traditional AI IDE"
        E1["Editor (Core)"] --> C1["AI Chat (Auxiliary)"]
    end
    subgraph "Antigravity"
        AM["Agent Manager<br/>Mission Control (Core)"] --> EV["Editor View"]
        AM --> BS["Browser Subagent"]
        AM --> AF["Artifacts"]
        AM --> SC["Source Control"]
    end

Traditional AI IDEs place the editor at the core, with AI chat as an auxiliary. Antigravity reverses this. Editor View and Agent Manager exist as peers, with Agent Manager serving as the product's core².

VS Code-based compatibility is maintained, and extension/keybinding migration is supported, keeping migration friction low. The essential difference is that the product's starting point is where you supervise agents, not where you write code.

Manager View: Operations Overview

UI Composition

According to the official codelab, the Manager screen's left navigation includes Start Conversation, Workspaces, Playground, Browser, and Editor View². This is a dashboard-style interface capable of simultaneously supervising dozens of agents across multiple workspaces.

The central area streams agent progress and pending approval tasks³. Approvals come in three types: terminal execution, browser operations, and implementation plans. The Inbox aggregates requests from all agents, allowing reviewers to confirm, approve, or deny multiple parallel tasks from one location.

Cross-Workspace Coordination

Automatic result routing between workspaces is not documented. The current approach is human-driven orchestration through files, Artifacts, comments, and global Workflows⁴.

In practice, you'd have an agent create an API spec in Workspace A, then reference those files from Workspace B's frontend agent.

Among the agents supervised by Manager View, the most distinctive is Browser Subagent.

Browser Subagent: Authentication and DOM Interaction

Separate Chrome Profile Design

The Browser Subagent operates in a Chrome profile separate from your regular Chrome². Everyday browsing cookies and sign-in info are not shared with the agent. However, sign-in state within the Antigravity-side profile is preserved and carries over to subsequent sessions.

For SaaS testing requiring login, log in once within the Antigravity-managed profile, and the agent can operate authenticated going forward.

Operational Model and Capabilities

The Browser Subagent runs a separate LLM from the main coding agent. DOM operations are handled by the Gemini 2.5 Pro UI Checkpoint model². Capabilities are as follows⁵:

Category	Specific Capabilities
Interaction	Click, scroll, text input
Information Retrieval	DOM capture, console log reading, screenshots
Analysis	Structured data extraction via Markdown parsing
Recording	Video recording of operations (auto-saved to Artifacts)

For Captcha bypass, iFrame limitations, and SSO-specific flow support, see the "Unconfirmed Items" section at the end.

Browser Subagent operations are automatically recorded in the Artifacts system described next.

Artifacts: Evidence-Driven Review Workflow

When an agent performs work, four types of data are automatically structured as Artifacts¹:

• Task lists (overall implementation plan)
• Implementation plans (records of judgment criteria)
• Screenshots (visual state snapshots)
• Browser recordings (Browser Subagent operation footage)

These can receive inline comments like Google Docs, and the agent adjusts its work based on feedback. In other words, plan → judgment basis → execution results → verification evidence are automatically packaged, eliminating the need for reviewers to manually dig through chat logs.

This mechanism is particularly valuable in enterprise environments requiring audit compliance, or when explaining code rationale to non-engineers.

Since Artifacts provide evidence tracking, the next critical concern is how to control agent permissions.

Governance UI and Security Settings

The more authority you grant agents, the more autonomy increases — but so does accident risk. Antigravity addresses this with a governance UI controllable from the settings screen.

Key Setting Items

Settings confirmed through official codelab and documentation²⁶:

Category	Setting Name	Purpose
Terminal	Terminal Command Auto Execution	Automation level for command execution
Terminal	Allow list / Deny list	Define permitted/prohibited commands
Terminal	Enable Terminal Sandboxing	Enable sandboxing
Terminal	Sandbox Network Access	Network permissions within sandbox
Browser	Browser URL Allowlist / Denylist	Restrict agent access targets
Browser	Browser JavaScript Execution	JS execution approval policy
Artifact	Artifact Review Policy	Review requirement settings
Workspace	Workspace external file access	Restrict references outside current workspace
General	Enable Telemetry	Telemetry data transmission permission

Strict Mode: Maximum Constraint

In Strict Mode, terminal execution is locked to Request Review, browser JS execution also requires review. Workspace external file access is prohibited, sandboxing is enabled, and network access denied⁷. This is multi-layered defense combining workspace isolation and network disconnection.

However, the existence of a governance UI and actual safety are separate matters. Community reports include D: drive deletion incidents and RCE vulnerability disclosures¹. Caution is warranted for sensitive projects during the preview stage.

Sandbox implementation details, Interrupt button names, and other items without official documentation are consolidated in the "Unconfirmed Items" section at the end.

With the governance UI defining "what to restrict," Agent Skills let you specialize agent capabilities for specific tasks.

Agent Skills: Activation Principles and Practical Operations

Skill Selection via Progressive Disclosure

Agent Skills are defined as folder structures centered on SKILL.md. Agents automatically select and activate them based on context.

The activation principle uses Progressive Disclosure⁸. At conversation start, agents read only each skill's name and description. The full body loads only when a request matches. Agents decide whether to use a skill even without explicit naming.

Placement and Conflict Avoidance

Skills can be placed in two locations:

Location	Path	Scope
Project-specific	<workspace>/.agent/skills/	Current workspace only
Global	~/.gemini/antigravity/skills/	Available across all workspaces

To avoid conflicts, design global skills broadly and workspace skills narrowly. Write specific trigger conditions AND "do not use" conditions in description. Overly broad descriptions like "help with code" cause misfires.

For detailed skill creation methods, see the Antigravity Skills Guide.

Setup and Initial Configuration

Installation and Model Selection

1. Download the OS-specific installer from the official site
2. Sign in with a Google account (no waitlist; available with personal Gmail)
3. Enable Enable Agent Mode from settings

Model selection is available by use case: Gemini 3 Pro for fast everyday coding, Gemini 3 Deep Think or Claude Opus 4.5 for complex design tasks.

Source Control Integration

From both Manager and Editor, Review Changes + Source Control enables viewing changed files, stage / unstage / commit upstream⁹. MCP enables external tool integration, and Workflows can define repetitive tasks.

For end-to-end GitHub PR creation through comment response, auto-fix, and re-push: deep automation is possible through Source Control + MCP + Workflow combinations, but native integration boundaries are covered in the "Unconfirmed Items" section.

Troubleshooting and Limitations

Stopping and Controlling Agents

Three stop pathways are confirmed³¹⁰:

• Inbox approval controls: Deny approval requests
• Chrome extension cancel: Cancel current conversation from the browser side
• Conversation deletion: Right-click delete in Agent Manager, or trash icon in Editor

Dedicated Interrupt button names and official procedures for infinite loops are covered in the "Unconfirmed Items" section.

Rate Limits and Plans

The free preview tier has rate limits. Specific values are not published, but reports indicate they've become stricter since initial release. Google AI Pro ($20/month) offers relaxation. Pricing changes are expected at GA, so current cost alone should not drive decisions.

Enterprise Adoption

The official codelab positions Antigravity as a personal Gmail account preview². Current plan info shows organizational tiers and BYOK/BYOE expressions, indicating enterprise expansion is in progress. For SOC 2 certification status, see the "Unconfirmed Items" section.

Conclusion

Antigravity's hallmark is a design where management → evidence → verification → review are integrated as a single operational unit. Individual features can be achieved with other tools, but having them built in from the start is the structural difference from other AI IDEs.

Preview-stage challenges remain. However, in environments requiring accountability and audit compliance for agent output, these integrated pathways can become the reason to choose it. If daily speed or CI/CD integration takes priority, CLI tools or Cursor may better fit current needs. Which operational pathways your environment requires is the essence of AI IDE selection.

Unconfirmed Items

Items covered in this guide without explicit official documentation, where hands-on verification is recommended.

Category	Unconfirmed Item	Notes
Browser Subagent	Captcha bypass capability	Not documented5
Browser Subagent	iFrame interaction limitations	Not documented5
Browser Subagent	SSO-specific flow support	Not documented5
Governance	Sandbox implementation basis (Docker / VM / seccomp)	Only "what is restricted" is documented, not implementation11
Troubleshooting	Dedicated Interrupt button name in Manager View	Not documented
Troubleshooting	Official procedure for infinite repair loops	cancel / delete / deny is the current practical approach
Git Integration	End-to-end PR creation → comment reply → auto-fix → re-push	Possible via SC+MCP+Workflow combo, but native boundary unverified
Enterprise	Antigravity-specific SOC 2 certification	Must be distinguished from general Google Cloud compliance
Context	Warning message at context limit	Not documented

Share this article Copy Link

Related Articles

• Google Antigravity In-Depth Review — Comprehensive feature review
• What Makes Antigravity Different — Structural analysis
• Antigravity Skills Guide — Complete beginner's guide to Skills

---

1. What Makes Antigravity Different — Structural analysis on this site ↩↩↩

2. Getting Started with Google Antigravity — Google Codelabs official tutorial ↩↩↩↩↩↩

3. Antigravity Inbox — Official documentation ↩↩

4. Antigravity Workspaces — Official documentation ↩

5. Browser Subagent — Official documentation ↩↩↩↩

6. Antigravity Settings — Official documentation ↩

7. Strict Mode — Official documentation ↩

8. Antigravity Skills — Official documentation ↩

9. Review Changes - Manager — Official documentation ↩

10. Chrome Extension — Official documentation ↩

11. Sandbox Mode — Official documentation ↩