Understanding Sub-agents

This article explores Sub-agents, one of the core features of Claude Code designed to improve productivity and scalability when working with AI-assisted development workflows.

As conversations grow longer, the context window—the AI's working memory—can become saturated, increasing the likelihood that earlier instructions and prompts will be forgotten. To address this limitation, we examine the agent delegation model, which isolates specific tasks into independent agent instances that can operate separately from the main session.

We will explore the four key characteristics that distinguish Sub-agents from the primary session: independent context, custom prompts, tool access control, and permission boundaries. The article also provides a practical guide to organizing Sub-agents through Markdown-based agent definitions and Frontmatter configuration, enabling scalable agent management at both the project and user levels.