Model Routing Role Definitions

Reference for choosing the right model_role when writing agent frontmatter or delegating to sub-agents. There are 13 roles organized into 5 categories.

Quick Decision Flowchart

What does your agent primarily DO?
│
├─ Writes code?
│   ├─ Frontend/UI code (components, layouts, CSS)? → ui-coding
│   ├─ Security vulnerability scanning? → security-audit
│   └─ General code (backend, IaC, tests, debugging)? → coding
│
├─ Thinks deeply / designs systems?
│   ├─ Designing architecture or solving complex problems? → reasoning
│   ├─ Reviewing/critiquing existing work for flaws? → critique
│   └─ Investigating and synthesizing from multiple sources? → research
│
├─ Creates content?
│   ├─ Aesthetic/design direction (visual, brand, style)? → creative
│   └─ Long-form writing (docs, marketing, case studies)? → writing
│
├─ Works with images?
│   ├─ Needs to understand/analyze images? → vision
│   └─ Needs to generate images? → image-gen
│
├─ Orchestrates high-stakes operations?
│   └─ Infrastructure, deployments, shadow environments? → critical-ops
│
├─ Quick utility task (parsing, classification, file ops)?
│   └─ → fast
│
└─ None of the above / genuinely varied work?
    └─ → general

Role-by-Role Reference

Foundation Roles

These two roles are required in every matrix and serve as universal fallbacks.

general

• Description: Versatile catch-all, no specialization needed
• Model tier: Mid (Sonnet, gpt-5.5, Gemini Pro)
• When to use: The agent's work is genuinely varied — knowledge experts, ecosystem consultants, integration specialists
• When NOT to use: If you can name what the agent primarily does (codes, designs, reviews), use that role instead
• Example agents: explorer, foundation-expert, amplifier-expert, core-expert
• Example chains: general (terminal — no fallback needed)