- →Agency Agents = a library of persona definitions. Tells the AI WHO to be.
- →Claude Sub-Agents = a built-in execution architecture. Defines HOW tasks are delegated in parallel.
- →They are complementary. Use them together for best results.
Agency Agents vs Claude Sub-Agents: What's the Difference?
The Core Confusion
Both terms contain the word “agent” and both relate to Claude Code. That's where the similarity ends. They solve completely different problems at different layers of the stack.
Side-by-Side Comparison
Agency Agents
Claude Sub-Agents
| Feature | Agency Agents | Claude Sub-Agents |
|---|---|---|
| Nature | Persona / prompt library | Execution architecture |
| Purpose | Shape AI behavior and identity | Delegate and parallelize tasks |
| Format | Markdown .md files | Native Claude Code feature |
| Tool support | 8 tools + growing | Claude Code CLI only |
| Open source? | ✅ MIT licensed | ❌ Closed/proprietary |
| Solves | Generic AI output quality | Task orchestration at scale |
| Use together? | ✅ Recommended | ✅ Recommended |
How They Work Together (The Right Pattern)
The most powerful workflow combines both. Here's how a senior developer might use them together:
Install the Security Engineer, Backend Architect, and Code Reviewer personas. Now each agent has a distinct identity and expertise.
Your lead Claude Code session spawns 3 sub-agents: one for security audit, one for architecture review, one for code review — all running in parallel with isolated context.
The Security Engineer sub-agent runs OWASP checks, the Backend Architect sub-agent reviews scalability, the Code Reviewer flags style issues. Three expert reviews in the time it takes for one.
When to Use Each
Use Agency Agents when...
- ✓You want consistent, expert-level output
- ✓You use Cursor, Copilot, or other tools
- ✓You need specific domain knowledge injected
- ✓You want to reduce AI hallucinations on specialized tasks
- ✓You want cross-tool consistency
Use Claude Sub-Agents when...
- ✓You need to parallelize work across multiple contexts
- ✓You want tasks to run without polluting the main context
- ✓You're orchestrating complex multi-step workflows
- ✓You need hub-and-spoke task delegation
- ✓You're using Claude Code exclusively
FAQ
Can I use Agency Agents and Claude sub-agents at the same time?
Yes, and this is the recommended approach. Use Agency Agents to define the persona (who the AI is and how it behaves), and use Claude Code's native sub-agent architecture to spawn and orchestrate multiple agents working in parallel. They are complementary, not competing.
Do I need Agency Agents if I just use Claude Code's native sub-agents?
Native Claude sub-agents give you the execution architecture — parallel task spawning, isolated context windows, delegation. But without persona definitions, each sub-agent is still a generic assistant. Agency Agents gives each sub-agent a specific expert identity, dramatically improving output quality.
Are Agency Agents specific to Claude Code?
No. Agency Agents personas work with Claude Code, Cursor, GitHub Copilot, Codex, Gemini CLI, Qwen Code, opencode, and Osaurus. Claude Code's native sub-agents, however, are specific to the Claude Code CLI.