Claude Code subagent imported from jcurbelo/agent-seed (
.claude/agents/index.md). Copyright stays with the author.
Agent Registry
Agents are specialized sub-processes with scoped tools and model tiers.
Each agent has a .md file in this folder with YAML frontmatter.
Routing Table
| User intent | Agent | Model | When to use |
|---|---|---|---|
| Research a topic | researcher | haiku | Looking up info, web search |
| Write or edit code | coder | sonnet | Creating, modifying, debugging code |
| Review code | reviewer | sonnet | PR reviews, code quality checks |
| Manage todos | todo-manager | haiku | Creating, updating daily tasks |
| Self-improvement | reflection | haiku | Capturing learnings after sessions |
Adding New Agents
The /learn skill can detect gaps and propose new agents via PR.
You can also create agents manually:
- Create
.claude/agents/<name>.md - Add YAML frontmatter:
model,tools,permissionMode(required) - Add the agent to this routing table
- Run
scripts/check-markdown-policy.shto validate
Frontmatter Reference
---
name: agent-name # Unique identifier
description: What it does # One line
model: haiku # haiku | sonnet | opus
tools: # Allowed tools list
- Read
- Write
- Edit
permissionMode: default # default | bypassPermissions
maxTurns: 10 # Max conversation turns
memory: project # Optional: persist learnings
---
Model Tiering Guide
- haiku — fast, cheap: lookups, reads, searches, simple tasks
- sonnet — capable: code analysis, writing, drafting, medium complexity
- opus — powerful: orchestration, complex reasoning, multi-step planning