Agents Overview

Seventeen subagents ship in agents/. Each one is a single Markdown file with the canonical D-09 frontmatter contract and a free-form body.

What an agent is

A subagent is a one-shot, role-prompted invocation of the host runtime's "spawn a subordinate agent" capability (Claude's Task, Codex's equivalents, …). Agents do not run in a daemon; they spawn, execute, and exit.

Workflows orchestrate. Agents read, write, and reason. The orchestration graph is exactly one level deep — agents may not spawn other agents.

How to read this section

• Catalog — every agent with tier, tools, and what spawns it.
• Frontmatter Schema reference — the four-gate validation contract, forbidden fields, and tier enum.
• Concepts → Agents — the design rationale (why no model, why no hooks, what tier means).

Tier breakdown

• opus (6 agents) — np-planner, np-plan-checker, np-code-reviewer, np-security-auditor, np-eval-planner, np-framework-selector.
• sonnet (7 agents) — np-researcher, np-executor, np-verifier, np-code-fixer, np-ai-researcher, np-ui-researcher, np-domain-researcher.
• haiku (4 agents) — np-ui-checker, np-ui-auditor, np-eval-auditor, np-nyquist-auditor.

The mandatory-initial-read invariant

Every agent body starts with the same critical guard:

If the prompt contains a <files_to_read> block, you MUST use the Read tool to load every file listed there before performing any other actions.

Skipping this is the single most common cause of agents that "ignore the plan" — they reason from prior assumption instead of orchestrator-supplied context. The verifier and plan-checker will catch the symptom, but it is cheaper to honor the contract.