Anthropic Engineers Propose "Agent Skills" Concept Instead of Creating New AI Agents for Each Task

Key summary

Barry Zhang and Mahes Muraj, engineers from Anthropic, presented the concept of "Agent Skills" at the 2025 AI Engineering Summit in New York City. They pointed out that instead of creating multiple AI agents for specific tasks, we can create a single agent and add knowledge about how to perform tasks through "Skills." This concept uses folders to store manuals, procedures, and resources, with a SKILL.md file as the core to define the name, description, and method of operation, rather than explaining a new workflow each time. The development is divided into 5 steps: 1) Select tasks that are frequently repeated and have clear results, such as summarizing reports or reviewing articles; 2) Complete the task once and have the AI summarize the manual from the conversation; 3) Provide clear names and descriptions so that the AI can decide when to use the skill when encountering a matching problem (Progressive Disclosure); 4) Correct Skills based on actual errors through an evaluation system and version management; 5) Share Skills as a knowledge base for the team. This concept applies to Claude, as well as ChatGPT and Codex from OpenAI, with each platform potentially having different installation details.