What are Agent Skills and How Agents Use Them?

Agent Skills are a three-tier progressive disclosure system for AI agents, introduced via the SKILL.md open standard released by Anthropic in December 2025. Instead of loading all workflow instructions upfront (which could consume ~150,000 tokens), agents advertise available skills at startup (~100 tokens each), load full instructions only when needed, and fetch reference files or scripts on-demand. Skill routing is handled entirely by LLM reasoning — no embeddings or classifiers. Skills differ from tools: tools execute discrete actions, while skills inject specialized procedural knowledge into the agent's context. The post also covers how Skills and MCP are complementary layers, with MCP handling connectivity and Skills providing domain expertise and workflows.