Sub Agents

This lets you spin up specialized Claude instances with their own prompts, tool access, and model selection. You define agents as markdown files in .claude/agents with YAML frontmatter specifying what they can do. The interactive /agents menu is the easiest way to start, though you can also write the configs by hand. Worth using when you want a code reviewer that only reads files, a debugger with full edit access, or a fast search agent running on Haiku. The ability to restrict tools per agent is the real win here. You can chain agents together or resume them later by ID, and project-level agents override your global ones.