Claude/Code/Subagent
- Claude/Code's implementation of AI/Agent/Subagent: a Markdown file with YAML frontmatter, resolved from
.claude/agents/ for the project, ~/.claude/agents/ for every project, or a Claude/Code/Plugin's agents/ directory, in that precedence order. - Only
name and description are required. The optional surface is the widest of any host — tools, disallowedTools, model, permissionMode, maxTurns, skills, mcpServers, hooks, memory, background, effort, isolation (set to worktree for a temporary git worktree), color, and initialPrompt. That richness is exactly what makes a Claude agent definition lossy to port anywhere else. - Each subagent runs in its own context window with its own system prompt, tool access, and permissions, so the delegating conversation gets back a summary rather than the search results.
- Claude/Code/Plugin/agent-teams-claude-code packages a set of these as a unit.