Subagents in Gemini CLI - Gemini CLI now runs specialist subagents in your terminal
by•
Gemini CLI's new subagents feature lets the main agent delegate complex tasks to specialised agents, each with isolated context, custom tools, and scoped permissions.
For developers building or automating from the terminal.


Replies
Super interesting launch 👀
What it is: Gemini CLI subagents — a system that lets your main AI agent delegate tasks to specialized “expert” agents.
Problem → Solution: Complex workflows overload a single agent’s context and slow things down. Subagents solve this by splitting work into isolated, task-specific agents that return clean, summarized outputs.
What makes it different:
Instead of one overloaded AI, you get a coordinated team of agents working in parallel — each with its own tools, context, and instructions.
Key features:
Parallel execution of multiple subagents
Isolated context windows (no context pollution)
Custom subagents via simple Markdown configs
Built-in agents for codebase analysis, CLI help, and general tasks
Easy delegation using @agent syntax
Benefits:
Faster execution on complex tasks
Cleaner context → better outputs
Scalable workflows for dev teams
Who it’s for:
Developers, AI builders, and teams working on large codebases or multi-step workflows.
Use cases:
Codebase exploration & debugging
Parallel research or analysis tasks
Automated workflows with custom agents
Enforcing coding standards across projects
This feels like a shift from “using AI” → “managing AI teams.” 🚀
DiffSense
@agent @rohanrecommends Gemini cli is the best deal by far right now. Also when you consider what you can do with it by connecting it with github local runners. PR reviews, Sec audits, Code quality bots etc.
@agent @rohanrecommends For a solo builder tweaking a large codebase, what's one simple @agent config you'd recommend first to debug cross-file dependencies without the main agent getting overwhelmed?
Great! congrats on the launch :)
isolated context per subagent is the right call — curious how scoped permissions work in practice. can the main agent grant a subagent temporary shell access for a single task, or is it declared upfront per subagent definition?
the markdown configs are the part that actually scales. been building with ai agents for months — every orchestration setup ends up in someone's head or a private doc, outside version control and outside code review. agent definitions that live in the repo changes the discipline in ways that gui-based approaches never could.