Agents

Lean reference for the built-in agents.

Why these agents

• Clear separation of roles so you can pick the right depth quickly.
• Safer behavior by default through scoped skill allowlists.
• Faster handoffs across planning, implementation, review, and docs.

Agent Overview

Agent	Best For
@orchestrator	Multi-phase coordination
@planner	Read-only architecture/planning
@codebase	Feature implementation
@review	Security/performance/code quality
@docs	Documentation updates
@em-advisor	Engineering leadership guidance
@blogger	Blog/video/podcast drafts
@brutal-critic	Final content quality gate

Canonical source for exact allowlists and skill triggers: Skills Matrix.

Suggested Flow

@orchestrator (plan)
→ @codebase (implement)
→ @review (audit)
→ @docs (document)

Skill Usage Guardrails

• All built-in agents support the skill tool.
• Skills are loaded on demand (not eagerly).
• Use one relevant skill per phase by default; add another only for clear cross-domain dependencies.
• If stack/domain is unclear, clarify before loading.

Skill Permission Model

• Agents use permission.skill rules to restrict which skills can load.
• Recommended baseline is "*": "deny" with explicit per-skill allows.
• This enforces least privilege and keeps agent behavior domain-scoped.

Task Permission Model

• Agents with Task access should use permission.task with deny-by-default and explicit allows.
• Keep subagent invocation scoped to role-appropriate handoffs.
• Rules are matched in order; last matching rule wins.

Skill Scope Policy

• Keep current core-only skill scope.
• Add skills only with repeat demand, clear gap, owner, and clean licensing/provenance.

Next Steps

• Commands & Skills
• Coding Standards
• Customization