Build an evaluator for the agent-skill boundary

[diberry]

Summary

This post explains how to build an evaluator that decides whether work belongs in a skill, an agent, or a hybrid pattern. It uses the same task under multiple candidate designs and scores each one against evidence, maintainability, safety, and ownership.

What this adds

• A practical mental model for the agent-skill boundary
• A scoring framework and design rubric
• Examples of skill-only, agent-only, and hybrid patterns
• Guidance on local vs MCP/connector execution and why they are separate concerns
• Plain-language explanations with diagrams for easier adoption

Why it matters

Many AI projects fail because organizations treat the words "agent" and "skill" as interchangeable. This post uses a neutral evaluator framing to make the decision explicit and evidence-based.

Validation

• Docusaurus build passed locally using the repo's standard build command
• SVG assets were added for the article and rendered correctly during the build