Google Adds Subagents for Parallel Coding to Gemini CLI

AI coding agents can take on complex tasks, but they usually handle them one at a time, which easily creates bottlenecks. Google has introduced subagents to Gemini CLI, allowing multiple specialized agents to divide roles and work within the same environment.

Each subagent has its own unique instructions, tools, and context. When the main agent delegates part of a task, the subagent processes it in a separate context and returns the result. This makes it possible to parallelize work, such as distributing frontend, testing, and documentation tasks to different agents after backend work is done.

The key point is that tasks are split and executed simultaneously within a single session. Since each subagent operates in an independent working space, instructions and outputs don't get mixed up, and it also reduces the problem of long conversations accumulating and becoming complex within one session.

The main usage methods are as follows.

• Using built-in subagents: generalist, CLI-only agent, codebase-focused agent
• Creating custom subagents: defined in a Markdown file with YAML frontmatter and plain-text instructions
• Direct assignment execution: assigning tasks to specific subagents using @ syntax
• Checking currently available agents: the /agents command

Google designed it especially so that repetitive or clearly role-defined tasks can be handed off to subagents. For example, a frontend-only agent could be launched simultaneously across multiple packages to analyze different areas of code respectively. Claude Code already supports a similar model, but Gemini CLI provides this tailored to division of labor and parallelization within a single session.