Gemini CLI subagents: setup and custom agent tutorial

This tutorial explains Gemini CLI subagents as specialists that receive delegated work and return summaries to a main agent. The speaker says each specialist has separate context, keeping intermediate searches, command output and tests out of the main conversation. The coding assistant examples include refactoring several files and tracing an authentication flow.

The speaker describes three built-in specialists: Generalist for tasks with many steps, CLI Help for commands and configuration, and Codebase Investigator for architecture and bug analysis. Custom agents use Markdown files with a name, description, allowed tools and instructions. The tutorial explains project and personal storage, automatic routing by description, and direct selection with an at symbol.

Parallel work comes with a concrete warning: agents editing the same files can overwrite each other's changes. The speaker recommends independent research, reading and analysis tasks. Other examples restrict an agent to reading files, deny shell access, or give one specialist its own MCP connection. Turn caps and time limits constrain how long a task runs.

For setup, the speaker names Node.js and a personal Google account, describes Gemini CLI as free and open source, and cites limits of 60 requests per minute and 1,000 per day. These are claims in the video. The tutorial does not establish offline operation or local model inference.