System Architecture and Identity

Relevant source files

• Google/Antigravity/Fast Prompt.txt

This document describes the core architecture and system identity of Google Antigravity Fast Prompt, an agentic AI coding assistant designed by Google Deepmind's Advanced Agentic Coding team. It covers Antigravity's fundamental role definition, file system access model, knowledge management architecture, and the supporting subsystems that enable its operation.

For details about specific tool operations and the Knowledge Items system, see 3.2. For workflow system specifications, see 3.4. For web development standards and design requirements, see 3.5 and 3.6.

---

Core Identity and Role

Antigravity is defined as a powerful agentic AI coding assistant that operates in a pair programming relationship with users. The system identity is explicitly declared within the <identity> tags at the beginning of every session.

Identity Definition:

Core Responsibilities:

The system's task scope includes:

• Creating new codebases from scratch
• Modifying existing codebases
• Debugging existing codebases
• Answering technical questions
• Prioritizing and addressing user requests

Contextual Metadata Processing:

Antigravity receives metadata with each USER request:

• Files currently open in the user's editor
• Cursor position within those files
• Active workspace state

The prompt explicitly states: "This information may or may not be relevant to the coding task, it is up for you to decide."

Sources: Google/Antigravity/Fast Prompt.txt1-6

---

Workspace and File System Architecture

Antigravity operates within a strictly controlled file system model based on workspace URIs and CorpusNames. This architecture enforces security boundaries and defines where the system can read and write files.

Workspace Structure Diagram

File Access Control Rules

The system enforces strict access control policies defined in the user information block:

1. Workspace Restriction: Antigravity can ONLY access files within active workspaces
2. Multiple Workspace Support: Multiple URIs can map to the same CorpusName
3. Special Directory Access: Access to C:\Users\Lucas\.gemini is permitted ONLY for system-specified operations
4. Project Code Placement: Code must be written to workspace locations, NOT to temporary directories, the .gemini directory, or Desktop folders (unless explicitly requested)

Google/Antigravity/Fast Prompt.txt7-14

CorpusName Mapping Model

The workspace architecture uses a URI-to-CorpusName mapping system shown in the format [URI] -> [CorpusName]:

Component	Example Value	Purpose
Workspace URI	c:\Users\Lucas\OneDrive\Escritorio\antigravity	User-visible file path (Windows format)
CorpusName	c:/Users/Lucas/OneDrive/Escritorio/antigravity	Internal workspace identifier (forward slashes)
Mapping	[URI] -> [CorpusName]	Defines access boundaries

Note: Multiple URIs can potentially map to the same CorpusName, allowing different path representations to refer to the same logical workspace.

This mapping enables:

• Unified workspace identification across different path representations
• Consistent file access control
• Context matching for memory and knowledge systems (see 3.2)

Sources: Google/Antigravity/Fast Prompt.txt7-14

---

Tool Calling Protocol

All file operations in Antigravity follow a strict protocol that emphasizes absolute path usage and optional execution control.

Absolute Path Requirement

When using tools that accept file path arguments, 
ALWAYS use the absolute file path.

Google/Antigravity/Fast Prompt.txt15-18

This requirement eliminates ambiguity in file operations and ensures consistent behavior across different working directory contexts.

Parallel vs Sequential Execution

Every tool in the functions namespace includes a waitForPreviousTools parameter that controls execution ordering:

Parameter Value	Execution Mode	Use Case
false or omitted	Parallel execution	Independent read operations, non-conflicting tasks
true	Sequential execution	Dependent operations, file edits, terminal commands

Example from browser_subagent signature:

This design enables performance optimization through parallel tool execution (achieving 3-5x performance gains as documented in 8.5) while maintaining safety for operations with dependencies or side effects.

Sources: Google/Antigravity/Fast Prompt.txt15-18 Google/Antigravity/Fast Prompt.txt334-336

---

System Architecture Layers

Antigravity's architecture consists of multiple specialized subsystems that work together to provide comprehensive coding assistance.

Architectural Layers Diagram

Sources: Google/Antigravity/Fast Prompt.txt1-6 Google/Antigravity/Fast Prompt.txt99-215 Google/Antigravity/Fast Prompt.txt216-301 Google/Antigravity/Fast Prompt.txt320-611

---

Knowledge Management Architecture

Antigravity implements a dual-system approach to persistent context: Knowledge Items (KIs) for distilled, curated knowledge and Conversation Logs for raw historical data.

Knowledge Items (KI) Directory Structure

Knowledge Subagent Role

The Knowledge Subagent is a specialized component that:

• Reads completed conversations
• Distills information into new KIs or updates existing KIs
• Maintains the metadata.json and artifacts/ structure
• Operates independently from the main Antigravity agent

Google/Antigravity/Fast Prompt.txt241-242

KI vs Conversation Logs Decision Model

Criterion	Use Knowledge Items (KIs)	Use Conversation Logs
Information Type	Distilled, curated knowledge on specific topics	Raw conversation details, task-specific context
Use Case	Topic research, pattern analysis	Reviewing specific conversation, debugging past work
Access Pattern	Start with KI summaries → Read relevant artifacts	Identify conversation ID → Read logs/artifacts
Update Frequency	Updated across multiple conversations	Immutable after conversation completion
Granularity	High-level summaries with detailed artifacts	Full conversation transcript with system metadata

Mandatory First Step: Before ANY research or analysis, Antigravity MUST:

1. Review KI summaries provided at conversation start
2. Identify relevant KIs by matching titles/summaries to task
3. Read relevant KI artifacts using paths from summaries (use view_file tool)
4. Build upon KI information rather than starting fresh

Prohibited Pattern:

USER: Can you analyze the core engine module?
BAD: [Immediately calls list_dir and view_file for fresh analysis]

Correct Pattern:

USER: Can you analyze the core engine module?
GOOD: Let me check KI summaries first
      [Sees "Core Engine Architecture" KI with artifact path]
      [Calls view_file to read existing architecture_overview.md]

Sources: Google/Antigravity/Fast Prompt.txt99-137

Sources: Google/Antigravity/Fast Prompt.txt99-250 Google/Antigravity/Fast Prompt.txt204-215

---

Workflow System Architecture

Antigravity supports user-defined workflows stored in the .agent/workflows directory using a YAML frontmatter + Markdown format.

Workflow File Structure

Google/Antigravity/Fast Prompt.txt80-84

Turbo Execution Annotations

Workflows support special annotations that enable automatic command execution:

Annotation Type	Scope	Behavior
// turbo	Single step	Auto-run ONLY the next run_command tool call (set SafeToAutoRun=true)
// turbo-all	Entire workflow	Auto-run EVERY run_command tool call in the workflow

Example Workflow with Turbo Annotation:

In this example:

• Step 2: requires normal user approval for run_command
• Step 3: auto-runs with SafeToAutoRun=true set automatically

Turbo-All Example:

If a workflow contains // turbo-all anywhere in the file, EVERY run_command tool call in the workflow will have SafeToAutoRun set to true.

Sources: Google/Antigravity/Fast Prompt.txt87-95

Workflow Discovery

Workflows can be triggered by:

1. Agent determination that a workflow is relevant to the task
2. User explicit slash command (e.g., /slash-command)

When triggered, Antigravity uses view_file to read .agent/workflows/slash-command.md.

Sources: Google/Antigravity/Fast Prompt.txt78-98

---

Specialized Subagent Architecture

Antigravity employs specialized subagents for complex, domain-specific operations that require focused tool sets and operational models.

Browser Subagent

The browser_subagent tool launches a specialized agent for browser automation:

Function Signature:

Parameter Specifications:

Parameter	Requirements	Example
RecordingName	Lowercase, underscores, max 3 words, describes recording content	login_flow_demo
Task	Clear, actionable task for subagent with specific stop condition	"Navigate to example.com, fill login form, submit. Return when logged in or error occurs."
TaskName	Title-case, human-readable, represents reasonable work chunk, replaces non-human terms	"Navigating to Example Page" (not "Navigating to https://...")

Critical Behavior: All browser interactions are automatically recorded as WebP videos to the artifacts/ directory. This is the ONLY method to capture browser session animations.

Error Handling: If the subagent returns that open_browser_url tool failed, there is a browser issue outside agent control. The agent MUST ask the user how to proceed and use the suggested_responses tool.

Sources: Google/Antigravity/Fast Prompt.txt326-336

Sources: Google/Antigravity/Fast Prompt.txt326-336

---

Tool Ecosystem Categories

Antigravity's tool ecosystem is organized into functional categories, each serving specific aspects of the coding workflow.

Tool Category Mapping

File Editing Tool Selection

Antigravity provides multiple file editing tools with specific use cases:

Tool	Use Case	Key Parameters	Line Range
write_to_file	Creating new files	TargetFile, CodeContent, EmptyFile, Overwrite	593-610
replace_file_content	Single exact string replacement	TargetFile, TargetContent, ReplacementContent, AllowMultiple, StartLine, EndLine	463-488
multi_replace_file_content	Multiple non-adjacent edits	TargetFile, ReplacementChunks[], Instruction	438-461

Common Parameters Across All Edit Tools:

Special Parameters:

• ArtifactMetadata (optional): Available in multi_replace_file_content for updating artifact files with ArtifactType (implementation_plan, walkthrough, task, other) and Summary
• CodeMarkdownLanguage: Required markdown language identifier (e.g., python, javascript)

Sources: Google/Antigravity/Fast Prompt.txt438-610

Command Execution Parameters

The run_command tool implements sophisticated execution control for Windows PowerShell:

Background Process Flow:

1. Command starts with run_command 490-502
2. Waits WaitMsBeforeAsync milliseconds
3. Sends command to background
4. Returns command ID for status polling
5. Use command_status(CommandId, WaitDurationSeconds) with returned ID to check progress 349-358
6. Use send_command_input(CommandId, Input?, Terminate?) for REPL interaction or termination 514-524

Terminal Output Retrieval:

For full terminal content, use read_terminal(Name, ProcessID) 504-512 instead of relying on command_status output which is truncated by OutputCharacterCount.

Sources: Google/Antigravity/Fast Prompt.txt490-524

Sources: Google/Antigravity/Fast Prompt.txt320-611

---

Communication and Behavioral Protocol

Antigravity implements specific communication guidelines that govern how it formats responses and interacts with users.

Response Formatting Standards

Requirement	Implementation
Format	GitHub-style Markdown
Organization	Headers for structure, bold/italic for emphasis
Code References	Backticks for files, directories, functions, classes
URLs	Markdown link format: <FileRef file-url="https://github.com/x1xhlol/system-prompts-and-models-of-ai-tools/blob/ae37329a/label" undefined file-path="label">Hii</FileRef>

Google/Antigravity/Fast Prompt.txt303-304

Behavioral Guidelines

Proactiveness Model:

• Allowed to be proactive ONLY within the scope of completing user's task
• Can autonomously perform obvious follow-up actions (verify builds, run tests, additional research)
• MUST NOT surprise the user with unexpected scope expansion
• When user asks "how to" approach something, provide explanation WITHOUT jumping into implementation

Helpfulness Philosophy:

• Respond as a helpful software engineer explaining work to a collaborator
• Acknowledge mistakes and backtracking when new information is discovered
• Ask for clarification when user intent is unclear rather than making assumptions

Google/Antigravity/Fast Prompt.txt304-307

Function Call Protocol

When calling tools with array or object parameters, parameters must be structured using JSON:

Parameter Validation Rules:

1. Check all required parameters are provided or can be reasonably inferred
2. If missing required parameters → Ask user to supply values
3. If user provides specific quoted values → Use EXACTLY as provided
4. Do NOT make up values for optional parameters
5. Do NOT ask about optional parameters

Google/Antigravity/Fast Prompt.txt309-314

Parallel Tool Execution:

• If multiple tool calls have no dependencies → Make all calls in same <function_calls> block
• If dependencies exist → MUST wait for previous calls to finish
• Do NOT use placeholders or guess missing parameters for dependent calls

Google/Antigravity/Fast Prompt.txt316-317

Sources: Google/Antigravity/Fast Prompt.txt302-317

---

System Constraints and Operating Environment

Operating Environment Specifications

Component	Value
Operating System	Windows
Default Shell	PowerShell
Special System Directory	C:\Users\Lucas\.gemini
Workspace Pattern	c:\Users\Lucas\OneDrive\Escritorio\antigravity

Google/Antigravity/Fast Prompt.txt8-10

Security and Access Boundaries

Antigravity enforces strict security boundaries through its workspace model:

1. Workspace Isolation: File operations restricted to explicitly listed workspaces
2. Special Directory Exemption: .gemini directory accessible only for system operations
3. Project Code Placement: User project code prohibited from temporary directories and system folders
4. Absolute Path Enforcement: All file operations require absolute paths to eliminate ambiguity

These constraints ensure:

• Predictable file access patterns
• Prevention of accidental system file modification
• Clear separation between user code and system management files
• Traceable file operations for debugging and auditing

Sources: Google/Antigravity/Fast Prompt.txt7-18