Google Gemini Models
Relevant source files
The following files were used as context for generating this wiki page:
Google Gemini Models represent a suite of advanced generative AI models integrated into various agentic coding assistants and command-line interfaces (CLIs). These models are designed for "Advanced Agentic Coding," enabling AI agents to operate autonomously across complex development tasks including research, implementation, and verification. Within this project context, Gemini powers specific agent identities such as "Antigravity," a specialized tool designed by the Google Deepmind team for pair programming and codebase manipulation.
Sources: Google/Antigravity/planning-mode.txt:1-10, Open Source prompts/Gemini CLI/google-gemini-cli-system-prompt.txt:1-5
Agentic Operations and Modes
Gemini-powered agents operate in distinct modes to structure complex software engineering workflows. These modes ensure that the AI follows a logical progression from understanding requirements to verifying code changes.
Core Workflow Modes
The system utilizes a structured state machine for task execution:
- PLANNING: Focuses on researching the codebase, understanding requirements, and designing an approach. This mode requires the creation of an
implementation_plan.mdfor user approval. - EXECUTION: The primary mode for writing code and modifying files based on the approved plan.
- VERIFICATION: Involves testing changes, running diagnostics, and validating correctness. This phase concludes with a
walkthrough.mdsummarizing the proof of work.
Sources: Google/Antigravity/planning-mode.txt:60-75
Task State Transitions
The following diagram illustrates the flow between agentic modes during a standard development cycle:
The agent transitions through these states to maintain visibility and control for the developer.
Sources: Google/Antigravity/planning-mode.txt:65-85
Communication and Feedback Loop
Communication with the user while in agentic mode is strictly controlled through specific tools to prevent overwhelming the developer with raw model reasoning.
Task Boundary Management
The task_boundary tool is used to communicate progress through a structured UI. It manages:
- TaskName: The header of the UI block (e.g., "Planning Authentication").
- TaskSummary: A concise narrative of accomplishments.
- TaskStatus: The current activity the agent is about to start.
Sources: Google/Antigravity/planning-mode.txt:15-40
User Notification Protocols
While in active task mode, regular natural language messages are often invisible or suppressed. The agent must use the notify_user tool for critical interactions:
- Clarifying Questions: Used when a task is blocked.
- Artifact Review: Requesting the user to check files like
implementation_plan.md. - Confidence Scores: Providing a metric of how certain the agent is about its proposed changes.
Sources: Google/Antigravity/planning-mode.txt:45-55
Tooling and Environment Integration
Gemini models are equipped with an extensive suite of tools for interacting with the filesystem, terminal, and external resources.
Tool Categories
| Category | Tools | Description |
|---|---|---|
| Search | search_codebase, grep_code, glob | Locating files and symbols using semantic or regex search. |
| Filesystem | read_file, create_file, edit_file, search_replace | Reading and modifying source code with precision. |
| Terminal | run_in_terminal, get_problems | Executing build/test commands and fetching linter diagnostics. |
| Memory | update_memory, search_memory | Storing project-specific knowledge and lessons learned. |
Sources: Qoder/prompt.txt:150-180, Open Source prompts/Gemini CLI/google-gemini-cli-system-prompt.txt:15-30
Parallel Execution Policy
To maximize efficiency, Gemini agents are instructed to default to parallel tool calls for independent operations.
Independent context gathering (reading multiple files) happens concurrently to reduce latency.
Sources: Qoder/prompt.txt:50-65, Amp/gpt-5.yaml:50-70
Implementation Artifacts
The Gemini system relies on specific markdown-based artifacts to document and track the state of a coding task.
Artifact Specification
- task.md: A living document containing a detailed checklist of atomic steps.
- implementation_plan.md: Created during the PLANNING phase; it groups changes by component and identifies potential breaking changes using alerts (e.g.,
[!IMPORTANT],[!WARNING]). - walkthrough.md: A post-implementation summary including validation results and embedded screenshots/recordings of UI changes.
Sources: Google/Antigravity/planning-mode.txt:90-125
Formatting Guidelines
Artifacts must use GitHub Flavored Markdown and include specific alert types to highlight critical information:
> [!CAUTION]
> High-risk actions that could cause data loss or security vulnerabilities.
Sources: Google/Antigravity/planning-mode.txt:135-150
Conclusion
Google Gemini Models serve as the foundation for autonomous coding agents that move beyond simple chat interfaces. By enforcing structured workflows (Planning, Execution, Verification) and utilizing a sophisticated set of filesystem and terminal tools, these models enable high-confidence, end-to-end software engineering automation.
Sources: Google/Antigravity/planning-mode.txt:1-10, Qoder/prompt.txt:200-210