Lore

Claude's Three Automation Layers: Skills vs. Sub-Agents vs. MCP

A three-way taxonomy for Claude's automation surfaces, each solving a different problem:

• Skills — always-active behavioral instructions (response styles, auto-detection patterns, design-system enforcement) that auto-detect and apply with no manual invocation, and persist across sessions.
• Sub-agents — complex, manually-invoked automation for multi-tool workflows, specialized processes, codebase access, state management, and decision trees.
• MCP (Model Context Protocol) — external-service connections ('one AI to another AI'), developed by Claude in November of the prior year; real-time, persistent, bidirectionally synced, but 'a bit sloppy' on security if not locked down. Likened to MIDI standardizing links between drum machines, synthesizers, and computers in a studio, or to a universal phone charger — the value is ecosystem standardization, not a new capability per se.

Practical implication: Sub-agents built for one-off Claude Code automations can later be converted into skill.md files (via Claude's built-in Skill Creator or a custom SOP-to-skill converter) so the same automation also runs in Claude Desktop.

Related: Claude Skills (Saving AI Analyses as Reusable Tools), MCP Response Throttling via Skills (Context-Window Preservation).

Из тем: Scaling, Omnichannel & Brand Growth