Gemini CLI launches sub-agents, allowing you to create a dedicated AI programming expert team using just Markdown files

ME News update: On April 16 (UTC+8), according to Dongcha Beating monitoring, Google has officially rolled out the subagents feature for its open-source command-line programming tool, Gemini CLI. Each subagent has its own independent context window, system instructions, and toolset. When the main agent handles complex tasks, it automatically delegates sub-tasks to the most suitable subagent; the execution results are compressed into summaries for return, without polluting the main session context.

Gemini CLI includes three built-in subagents: generalist (generalist agent, suitable for high tool-call volume tasks such as bulk refactoring), cli_help (an expert on Gemini CLI usage guides), and codebase_investigator (codebase exploration, architecture mapping, and bug tracing). Users can also explicitly specify which subagent should take over via the @agent syntax—e.g., @codebase_investigator for mapping the authentication workflow. The bar for creating custom subagents is very low: simply place a Markdown file with YAML front-matter in the .gemini/agents/ directory. In the file, define the role description, available tools, and system instructions. Files placed under ~/.gemini/agents take effect globally, while placing them in a project directory lets them be shared with the team alongside the code repository. Subagents can also be packaged into Gemini CLI extension distributions.

Subagents support parallel execution, and multiple subagents can simultaneously handle different sub-tasks. However, Google notes that parallel tasks involving code editing may create write conflicts, so caution is advised.

The day before, Anthropic also released a redesigned desktop version of Claude Code, whose core likewise centers on multi-agent parallelism. AI programming tools are collectively moving from a “single-agent conversation” model to an architecture of “main agent orchestration + expert agent execution,” driven primarily by context window management: as programming tasks become more complex, stuffing all intermediate steps into one context leads to performance degradation and higher costs—splitting them into multiple isolated agents is the current engineering consensus.

(Source: BlockBeats)