write-subagent

Create subagent configurations in the .claude/agents/ directory as YAML/Markdown files. Each subagent defines a single responsibility, isolated context window, minimal tool access, and trigger-based prompts. Define metadata including name, description, tools, model, skills, permissionMode, and invoke via the Task tool when needed.

Subagents enable delegation by isolating each task into dedicated agent configurations with specialized tools and clear triggers. A code-reviewer subagent and a testing subagent each operate in separate context windows, preventing context pollution and ensuring focused execution on their single responsibility.

Yes. Subagent configurations enforce isolated context windows for each agent. Multiple subagents can run independently without sharing context, enabling parallel execution of code review, testing, and domain-specific tasks without interference.

Each subagent configuration explicitly declares its tool access in the YAML metadata. Grant only necessary tools per subagent responsibility—a code reviewer may access file-reading tools; a tester may access build and test runners—minimizing scope and security surface.

Subagent metadata includes name, description, tools list, model selection, skills, permissionMode, and trigger-based prompts. This structured YAML/Markdown block in .claude/agents/ defines the subagent's identity, capabilities, and activation conditions.

Yes. Subagents support any domain-specific task with a clear, isolated responsibility. Define custom subagent configurations for API documentation generation, security audits, performance analysis, or other specialized workflows by declaring appropriate tools and triggers.