Gemini CLIのサブエージェントがタスク委譲と並列エージェントワークフローを可能に

Google has introduced subagents in Gemini CLI, a new capability designed to help developers delegate complex or repetitive tasks to specialized AI agents operating alongside a primary session.

The feature allows the main agent to act as an orchestrator, assigning subtasks such as code analysis, research, or testing to specialized subagents. Each subagent operates in an isolated environment and returns a summarized result to the main session, minimizing context overload and enhancing performance during longer interactions.

According to Google, this approach is intended to address common limitations in agent workflows, particularly the accumulation of intermediate steps that can slow down responses and increase costs. By offloading detailed operations to subagents, the primary agent remains focused on higher-level reasoning and final outputs.

Subagents can also run in parallel, enabling multiple tasks to be executed simultaneously. For example, developers can instruct the system to analyze different parts of a codebase or perform multiple research tasks at once. While this can reduce overall execution time, Google notes that parallel execution may introduce risks such as conflicting code changes and increased usage limits due to concurrent requests.

A significant aspect of this feature is customization. Developers can create their own subagents using Markdown files with YAML configuration, which allows them to define roles, tools, and behavioral guidelines. These agents can be saved locally or in a repository, enabling teams to standardize workflows or enforce coding practices across projects. Additionally, Google offers several built-in subagents, such as a general-purpose assistant, a command-line interface (CLI) helper, and a codebase investigation agent.

The system enables explicit delegation via prompt syntax, allowing users to assign tasks to specific agents directly. This provides developers with greater control over task distribution, rather than depending solely on automatic routing.

The release highlights a trend towards multi-agent architectures, where separate components manage specific tasks instead of relying on a single model, improving scalability and maintainability in complex development processes.

However, feedback from early users suggests that the overall developer experience still has room for improvement. One comment highlights ongoing concerns:

Google should invest more in stability and UI/UX for gemini-cli. At the moment, the experience, even with the Pro plan, is quite poor. The models are pretty good, but you should work a bit more on the tool set.

While the introduction of subagents expands the capabilities of Gemini CLI, adoption may depend on how quickly usability and reliability issues are addressed alongside feature development.

About the Author

Robert Krzaczyński

Robert Krzaczyński is a software engineer with solid experience in developing web applications. Passionate about applying artificial intelligence algorithms in medicine and the broader healthcare sector, he continuously expands his expertise in ML and AI. He holds a BSc Eng degree in Control Engineering and Robotics, as well as an MSc Eng degree in Computer Science.

Show moreShow less