## Overview Every major AI platform is independently converging on the same three-layer architecture for reliable agent workflows: **Skills** (reusable knowledge recipes), **Sub-agents** (parallel execution workers), and **MCPs** (system connectors/tool interfaces). This represents the first actual consensus on how to build reliable AI workflows — suggesting these three components are structurally necessary, not design choices. - **Key Principle**: Architectural convergence across competing platforms signals structural necessity rather than fashion — Skills, Sub-agents, and MCPs are the minimal viable agent architecture - **Connection**: Extends [[SKILL.md vs MCP Decision Framework]] (Skills vs MCPs distinction), [[Agent Harness Over Framework Pattern]] (harness as the orchestration layer), [[Agent-First Product Design]] (products must expose machine-readable interfaces) - Source: [[The architecture everyone's building toward just became obvious]] (Ishaan Sehgal, January 2026)