agent-authoring

Sub-Agent Authoring

Sub-agents run in a fresh context window with a restricted tool allowlist. Use them when:

• The task benefits from isolation (heavy file reading, research).
• You want deterministic tool scoping (read-only codebase analysis, no network).
• You need parallelism (spawn N agents concurrently).
• Context pressure in the orchestrator is high and you need to keep it lean.

Do not create a sub-agent for trivial work — the spawn cost (~1-3s, fresh context reload) exceeds benefit for tasks under ~5 tool calls.

Frontmatter schema (Claude Code plugin agents)

---
name: <kebab-case, matches filename>
description: What it does and when to invoke. Trigger-optimized like a skill description.
model: sonnet | opus | haiku    # optional, defaults to parent
effort: low | medium | high     # optional
maxTurns: 20                    # cap tool loops
tools: Read, Grep, Glob, Bash   # ALLOWLIST — omitting = inherit all
disallowedTools: Write, Edit    # alternative: blocklist
skills: [skill-a, skill-b]      # preload specific skills
memory: preserve | fresh        # context handling
background: true | false        # long-running background agent
isolation: worktree             # only valid value; copies repo to temp worktree
---

Security: hooks, mcpServers, and permissionMode are NOT allowed in plugin-shipped agents. Do not include them.

Body structure (XML-tagged, like GSD convention)

<role>
You are <specialist>. Spawned by <parent command/orchestrator>. You answer <single question> and produce <single artifact>.
</role>

<required_reading>
Files the orchestrator MUST pass via Read. List them.
</required_reading>

<responsibilities>
Bullet list, 3-7 items.
</responsibilities>

<constraints>
- Read-only / write-only / deterministic
- Never do X
- Always do Y
</constraints>

<output_contract>
Produce exactly: `<path>/<file>.md` with sections: ...
Exit with a 5-line summary for the orchestrator.
</output_contract>

Design principles

1. Single artifact: each agent produces one canonical file (PATTERNS.md, RESEARCH.md, PLAN.md). Never have an agent update multiple cross-cutting files.
2. Read-only by default: specialists read and analyze; only executor-class agents write code. Enforce via tools allowlist.
3. Tight tool allowlists: research agent → Read, Grep, Glob, WebFetch, WebSearch. Code reviewer → Read, Grep, Glob, Bash. Never include Write unless the agent's job is to write.
4. Explicit upstream contract: agents list exactly which files/variables they need; orchestrators pass them. No implicit shared state.
5. Exit summary: 3-7 lines summarizing what was produced + key findings. Keeps orchestrator context lean.

Orchestrator ↔ specialist pattern

The orchestrator (usually a slash command) routes. It:

1. Initializes state (reads configs, creates directories).
2. Spawns specialists via Task tool with subagent_type + prompt containing <required_reading> paths.
3. Aggregates their output files (reads frontmatter only, not full body — see context-management skill).
4. Makes the next routing decision.

The orchestrator NEVER does the specialist's work inline.

References

• @../../agents-architect/references/agent-tool-allowlists.md — canonical tool sets per agent archetype
• @../../agents-architect/templates/agent.template.md
• @../../agents-architect/references/context-budget.md — how orchestrators read agent output without blowing context

Output

Produce agents/<name>.md with correct frontmatter + structured body. Include a tool-allowlist justification comment for future auditors. Finish with a 3-line summary of the agent's niche.