Designs and authors Claude Code sub-agents (.claude/agents/*.md files) that integrate with an existing skill-stack. Use when: subagent, sub-agent, agent design, .claude/agents, custom agent, agent frontmatter, verifier agent, planner agent, agent description, agent delegation, Explore, Plan, general-purpose, Agent tool, agent permissions, agent tools field, agent skills preload, agent memory field, isolation worktree, permissionMode, agent decomposition, multi-agent decision, planner SPEC, verifier checklist, agent context isolation, tool-restricted agent, нужен ли мне агент, как сделать сабагента, как настроить .claude/agents. SKIP: building skills (→skill-evaluation), evaluating LLM agents in production (→agent-evaluation), Claude API / Messages SDK outside Claude Code (→claude-api), MCP server authoring (→mcp-server-author), agent-team coordination across separate sessions (→agent-teams), codex-specific agent authoring (use codex own conventions).
Tests and benchmarks LLM agents covering behavioral testing, capability assessment, reliability metrics, and production monitoring. Use when evaluating agent quality, designing eval suites, building regression tests, or measuring real-world reliability beyond benchmark scores.
System architecture discipline from Newman + Khononov + Richards/Ford. Use when: project-architect designs a new system, decomposes a monolith, chooses architecture style, writes ADRs, defines service boundaries. Trigger terms: bounded context, subdomain, architecture quantum, ADR, fitness function, saga, strangler fig. SKIP: single-file refactor, tactical UI design.