GitHub user wenjin272 created a discussion: Agent Skills Investigation
# Agent Skills Investigation ### Introduction Agent Skills are folders of instructions, scripts, and resources that agents can discover and use to do things more accurately and efficiently. #### example Take the `Brave Search Skill` as an example:：[https://clawhub.ai/steipete/brave-search](https://clawhub.ai/steipete/brave-search), It includes the following files:  * `SKILL.md` describes the skill’s name, purpose, and usage. * `search.js` and `content.js` are Node.js scripts used for web search and content extraction, respectively. ### Why Agent Skills * **For skill authors**: Build capabilities once and deploy them across multiple agent products. * **For compatible agents**: Support for skills lets end users give agents new capabilities out of the box. * **For teams and enterprises**: Capture organizational knowledge in portable, version-controlled packages. My understanding is: * Dynamically expand knowledge and capability boundaries—similar to RAG retrieving SOPs or domain-specific knowledge. * Increase determinism and reliability in task execution—akin to explicitly defining workflows. * Share skills across any agent that supports the agent skill—similar to how MCP servers work. ### How skills work Skills use progressive disclosure to manage context efficiently: 1. **Discovery**: At startup, agents load only the name and description of each available skill, just enough to know when it might be relevant. 2. **Activation**: When a task matches a skill’s description, the agent reads the full SKILL.md instructions into context. 3. **Execution**: The agent follows the instructions, optionally loading referenced files or executing bundled code as needed. In terms of implementation in agent framework: 1. During initialization, the agent loads the `name` and `description` of all skills from a specified directory and injects them into the system prompt. 2. A "load skill" tool is provided; when the LLM decides a skill is needed, it generates a tool call to fetch the skill’s full content. 3. A shell command execution tool is provided; when the LLM determines that a script or command from the skill should be run in the shell, it generates a tool call to execute it.  #### agentscope-java implementation We evaluated three repositories: LangChain, AgentScope, and AgentScope-Java. Among them, AgentScope-Java provides the most complete support for Agent Skills: [https://github.com/agentscope-ai/agentscope-java/issues/92](https://github.com/agentscope-ai/agentscope-java/issues/92). It mainly includes three components: * Implement basic skill registration, loading, and integration * Support sandboxed execution of scripts within skills * Improve ecosystem compatibility and maintainability GitHub link: https://github.com/apache/flink-agents/discussions/536 ---- This is an automatically sent email for [email protected]. To unsubscribe, please send an email to: [email protected]
