Agent skills are emerging as a new standard, and adoption is accelerating. OpenAI is quietly moving in this direction. Skills are already available in ChatGPT and Codex. Claude’s skills, in practice, often perform better where MCP falls short. They follow the path of least resistance. A skill is essentially a small, focused manual that an agent loads into its system context. The agent autonomously decides when to apply it based on the request. The key difference from MCP is subtle but important. Skills don’t introduce new tools. They reuse existing ones, like bash, and provide precise guidance on how to use those tools in the context of a specific problem or workflow. This will likely evolve over time. For now, manually maintained skills and agents that write their own tools have become my preferred approach.