Agentic AI just got a little more organized. Google announced this week that subagents have officially arrived in Gemini CLI, giving developers a structured way to break large, complex tasks into smaller, focused workloads handled by specialized agents running in parallel.
The feature, now enabled by default, marks a meaningful step forward for developers who use Gemini CLI as their primary AI-assisted coding environment.
What Subagents Actually Do
The core idea is straightforward. Subagents allow Gemini CLI to delegate complex, repetitive, or high-volume tasks to specialized expert agents. Each subagent operates within its own separate context window, custom system instructions, and curated set of tools.
Think of it like a project lead handing off work to subject matter experts. The main Gemini CLI session stays focused on the big picture while individual subagents handle specific jobs — a security audit here, a documentation lookup there, a deep codebase review somewhere else.
Each subagent has its own system prompt and persona, and interactions with a subagent happen in a separate context loop, which saves tokens in your main conversation history. That last part matters. Context bloat is one of the more frustrating friction points in working with long-running AI sessions. Subagents help keep the primary session lean.
How the Orchestration Works
Subagents are exposed to the main agent as a tool of the same name. When the main agent calls the tool, it delegates the task to the subagent. Once the subagent completes its task, it reports back to the main agent with its findings.
You can trigger subagents through automatic delegation or by explicitly calling them in your prompt. In automatic mode, Gemini CLI recognizes when a task fits a subagent's defined scope and routes it accordingly. If you prefer more control, you can call a specific subagent directly.
Custom subagents are defined as Markdown files with YAML frontmatter. Developers can place them in .gemini/agents/ at the project level or ~/.gemini/agents/ at the user level, making it easy to create reusable specialist agents tuned to specific codebases or workflows.
Remote subagents are also supported. Gemini CLI can delegate tasks to remote subagents using the Agent-to-Agent (A2A) protocol — opening the door to distributed agent architectures useful in enterprise environments where different teams or services own different parts of a development pipeline.
What to Watch Out For
There's a security consideration worth flagging. Subagents run in what Google calls YOLO mode by default, meaning they execute tools without prompting for user confirmation. For a subagent configured with file write or shell command access, that's a real risk if the agent is misconfigured. Teams adopting subagents in production workflows should restrict tool access carefully and test configurations in safe environments before going broad.
The v0.36.0 release introduced a multi-registry architecture and native macOS Seatbelt and Windows sandboxing for enhanced subagent security. That's a positive move, but sandboxing alone doesn't replace thoughtful configuration.
The Broader DevOps Angle
For DevOps teams, subagent architecture aligns well with how modern pipelines already work — distributed, modular, and specialized. The idea of an orchestrating agent spinning up specialist agents on demand fits naturally into CI/CD thinking. Instead of a single AI assistant trying to manage an entire software delivery workflow, you now have the option to assign discrete agents to discrete stages.
The practical applications are already taking shape. A developer could configure subagents for security scanning, test coverage analysis, dependency auditing, and documentation generation — each operating with purpose-built instructions and tool access. The main session ties the results together without having to carry the full weight of every intermediate step.
Getting Started
Subagents are enabled by default in current versions of Gemini CLI. Developers can explore the official documentation for configuration guidance, including how to restrict tools, write effective agent descriptions, and set up remote agents via A2A. The Gemini CLI GitHub repository is also a good reference for real-world examples.
For teams already using Gemini CLI in their development environments, this is a natural next step. It doesn't require a rethink of existing workflows — it just gives you more options for how to structure the work those workflows depend on.
