Explaining AI Agent, MCP, and Skills Using Household Management Metaphors

This article uses the metaphor of a household manager to explain the roles and operations of Agent, MCP, and Skills. The household manager represents the AI Agent, possessing foundational knowledge and the ability to plan and utilize tools. The robotic vacuum corresponds to the SubAgent, which is delegated and supervised by the main AI. The smart home protocol corresponds to the MCP, providing a unified access standard for tools. The 'Household Skills Manual' corresponds to Skills, containing a table of contents and detailed chapters. The table of contents serves as metadata (about 100 words), while the chapters provide specific instructions (less than 5000 words). Skills support 'dynamic loading' and 'progressive disclosure' to read on demand and save context window. 'My Home' corresponds to the Project, storing household manuals and historical records, with the manager writing work records. If a skill is untrained, the Agent will use search, self-made tools, or try other similar tools. The availability of dishwashing skills depends on system prompts allowing tools and whether the manual's table of contents includes it. Google will launch a five-day AI Agent course in November, with five white papers covering topics such as Agents, MCP, context, and quality, all available in PDF format.
Key insights
Household Metaphor: Correspondence Between Agent, MCP, and Skills
This long post uses the scenario of a household manager to systematically correspond AI components: the household manager represents the AI Agent, possessing foundational knowledge similar to a large language model and trained to plan and use tools; the robotic vacuum represents the SubAgent, a professional autonomous executor delegated and supervised by the main AI; the smart home protocol is labeled as MCP (Model Context Protocol), providing unified access for supported 'appliances/tools'; the 'Household Skills Manual' is labeled as Skills, containing a short table of contents and detailed chapters to supplement untrained skills of the manager. This metaphor illustrates the division of labor and interoperability of each part.
Dynamic Loading and Progressive Disclosure of Skills to Save Context
The post details the two-layer structure and reading strategy of Skills: first reading the table of contents (metadata, about 100 words) to determine relevance, then opening specific chapters on demand (complete instructions, <5000 words); this is termed 'dynamic loading' and 'progressive disclosure'. The direct purpose of this mechanism is to save 'mental/context window', ensuring the Agent utilizes key professional knowledge with minimal context when most needed. The post also points out that Projects (like household manuals and historical records) and work records support reference for the next task.
SubAgent and Tool Usage: Delegation, Supervision, and Execution Strategies
The post suggests that when the main Agent requires external execution, it will delegate to a SubAgent (like a robotic vacuum) to complete specific tasks, thus freeing the main Agent's context resources. If a skill is untrained, the Agent will use online searches to obtain guidance, create its own tools, or attempt to use other similar tools to accomplish tasks. The post also points out that tools can be allowed by system prompts or provided by the Skills manual, both determining whether the Agent prioritizes using a certain tool or consulting the skills manual.
Educational Resources: Google's November AI Agent Five-Day Course and Five White Papers
A post outlines the structure of Google's five-day AI Agent course launched in November, with five daily white papers covering the following topics: 1) Introduction to Agents; 2) Agent Tools & Interoperability with MCP; 3) Context Engineering: Sessions, Memory; 4) Agent Quality; 5) Prototype to Production. The post states that all five white papers are available in PDF format, serving as resources for further learning about Agents, MCP, and context engineering.