Google shipped subagent support for Gemini CLI on Wednesday, turning the terminal tool into something closer to an agent orchestrator. The announcement post pitches it as a fix for context rot: the main session stays lean while specialist agents handle heavier work in their own context windows.
Each subagent gets its own system prompt, tool set, and MCP server configuration. Three ship by default, including a general-purpose agent, a codebase investigator, and a helper for the CLI's own documentation. Developers can define custom ones through Markdown files with YAML frontmatter, dropped into .gemini/agents/ at the project level or in the user home directory. The subagent docs cover the config schema.
Parallel execution works too. Google flags its own caveat: multiple agents editing code simultaneously can cause conflicts and burn through usage limits faster. Remote subagents over the Agent-to-Agent protocol are also supported, with source on GitHub.
The approach echoes Anthropic's earlier subagents work for Claude Code. Subagents are on by default in Gemini CLI; flipping enableAgents to false in settings.json turns them off.
Bottom Line
Gemini CLI users can now spin up custom specialist agents via Markdown files in .gemini/agents/, with three built-in options active by default.
Quick Facts
- Announced April 15, 2026 on the Google Developers Blog
- Three built-in subagents: general-purpose, codebase investigator, cli_help
- Custom agents defined via Markdown files with YAML frontmatter
- Supports the Agent-to-Agent (A2A) protocol for remote subagents
- Enabled by default; toggled via enableAgents in settings.json
