Google Introduces Subagents in Gemini CLI

Google has rolled out subagents in Gemini CLI, a major enhancement that allows developers to delegate tasks to specialized agents running in isolated contexts. This innovation streamlines workflows, reduces context clutter, and enables parallel execution of multiple expert agents.

What Are Subagents?

• Subagents are specialized agents that operate alongside your main Gemini CLI session.
• The primary agent orchestrates tasks, while subagents handle subtasks independently.
• Each subagent has its own context window, system instructions, and toolset, ensuring isolation from the main session.

Key Benefits

• Improved efficiency by keeping the main agent focused on strategy and final responses.
• Faster execution through parallel processing.
• Cleaner context management, as subagents return summaries instead of overwhelming the main session.
• Flexibility to define custom subagents tailored to specific workflows.

Creating Custom Subagents

Developers can build their own expert agents:

• Defined in .md files with YAML frontmatter.
• Stored globally in ~/.gemini/agents or project-level in .gemini/agents.
• Packaged into Gemini CLI extensions via an agents/ directory.

For example, a frontend-specialist subagent could enforce accessibility, performance, and scalability standards in web applications.

Parallel Execution

Gemini CLI supports multiple subagents running simultaneously.
Example: “Run the frontend-specialist on each package in parallel.”
However, parallel edits can cause conflicts if multiple agents modify code at once, so best practice is to use parallel execution for research or analysis rather than heavy code edits.

Built-in Subagents

Gemini CLI includes several ready-to-use subagents:

• generalist: Handles broad, high-volume tasks.
• cli_help: Expert on Gemini CLI documentation.
• codebase_investigator: Specialized in exploring codebases, mapping architecture, and debugging.

How to Use Subagents

• Gemini CLI can automatically delegate tasks to subagents when efficient.
• Developers can explicitly delegate using @agent syntax in prompts.
  Examples:
• @frontend-specialist Review our app for improvements.
• @generalist Update license headers across the project.
• @codebase_investigator Map authentication flow.

Risks and Considerations

• Parallel subagents consume more requests, potentially hitting usage limits faster.
• Code conflicts may arise if multiple agents edit simultaneously.
• Recommended to use parallel execution for analysis rather than direct code changes.

Getting Started

• Run /agents in Gemini CLI to view configured subagents.
• Documentation provides details on configuration, tool restrictions, and optimization.
• Follow Gemini CLI updates on GitHub and X for new features.

---

Conclusion: Subagents transform Gemini CLI into a multi-agent orchestration platform, enabling developers to scale workflows, automate specialized tasks, and keep their main session lean. This marks a significant step toward making AI-powered development tools more modular, efficient, and collaborative.