Skills vs. Agents

I have 70+ AI agents created with my skills-creator in Claude Code. For a few weeks, I had no idea how to organize them.

Last week I wrote about precompiling context (what agents know). This is about structure (how they’re arranged).

The breakthrough was one distinction:

Skills vs. Agents.

In Claude, a skill is a persona you talk with. It loads context into your current session. It’s collaborative: back-and-forth, iterative.

An agent is a worker you dispatch. You send it a task. It returns a result. No conversation needed, just output.

Two modes. That’s the framework.

My “inner circle” skills are the ones I want to converse with (maybe 8—10): orchestrator, designer, implementer.

Everything else: protocol specialists, QA sub-agents, sync agents — are workers. They live in subdirectories, organized however makes them easy to find and easy to delegate.

70+ agents sounded unmanageable. Turns out only ~10 needed to be conversational. The rest just needed to be discoverable.

Not every agent needs a seat at the table. Most just need a job description.

The line is blurrier in practice: some workers eventually get “promoted” into skills when I realize I need iteration.

What’s the agent count where organization starts breaking down for you?