Andrew Ng x Anthropic: A Complete Guide to Agent Skills

DeepLearning.AI and Anthropic have jointly launched a hands-on course on Agent Skills, taught by instructor Ellie Sherwick. The course systematically covers how to extend Claude Code and other AI agents with new capabilities through "skills," quickly becoming one of the most talked-about introductory resources for agent development in the developer community.

Background: DeepLearning.AI was founded by Stanford professor and former head of Google Brain and Baidu AI Lab, Andrew Ng, focusing on AI education for engineers. It has previously co-launched several well-received courses with OpenAI, Google, Meta, and others. Anthropic is an AI safety company founded in 2021 by former OpenAI VP of Research Dario Amodei and colleagues. The Claude model family is its flagship product, and Claude Code is its AI coding assistant for developers, supporting direct terminal access to Claude for writing code, debugging, and project management.

This article draws on the core content of that course to break down how Agent Skills work, the design philosophy behind them, and their practical applications — helping developers understand this agent extension mechanism that is evolving into an open standard.

What Are Agent Skills

Skills are essentially a collection of instruction files organized into a folder, injecting specialized knowledge to expand an agent's capability boundaries. When you need an agent to repeatedly execute a particular workflow, instead of re-explaining it every time, you can package it as a skill — and the agent will automatically know how to handle it.

One of the most noteworthy points in the course is that Skills have become an Open Standard. This means they have a standardized format and can work with any Skills-compatible agent. Developers only need to build a skill once and deploy it across multiple agent products, completely eliminating the cost of reinventing the wheel.

The deeper significance of an open standard: The open standard positioning of Skills is analogous to the relationship between HTML/CSS and browsers in web development — the standard is defined by the community, independently implemented by vendors, and written once to run anywhere. In the AI agent space, the long-standing absence of open standards has been a root cause of ecosystem fragmentation: prompt formats, tool-calling conventions, and memory management approaches differ across platforms, making it hard for developers to migrate and reuse their work. The introduction of a standardized Skills format aims to turn skill files into universal assets across products, laying the groundwork for interoperability across the entire Agent ecosystem.

Every skill should include a skill.md Markdown file that defines the skill's name, description, and primary instructions. The primary instructions can also reference other files, such as scripts, supplementary documentation, and asset files like templates and images.

Progressive Disclosure: How Skills Are Loaded by an Agent

One of the most elegant design concepts in the course is Progressive Disclosure, which directly determines how efficiently the context window is utilized.

Context Window refers to the maximum number of tokens a large language model can process in a single inference pass — one of the most critical resource constraints in current LLM architectures. For example, Claude 3.5 Sonnet has a context window of approximately 200K tokens, which sounds generous but can be rapidly consumed in complex agent workflows by system prompts, tool definitions, conversation history, retrieved results, and more. Progressive Disclosure borrows from the UX design concept of the same name — only showing users relevant information when they need it — and applies it to LLM context management to dynamically balance capability breadth with inference efficiency. This is one of the central engineering challenges in productionizing agents today.

Layered Loading Mechanism

Skills are not all loaded into the agent's context at once. Instead, they are loaded on demand in layers:

• The skill's name and description always reside in the context window as a lightweight "index";
• The full instruction content of a skill is only loaded when a user's request matches the skill's description;
• Beyond that, the agent will further fetch referenced files and asset files if needed.

The value of this layered design is that an agent can "know" what capabilities it has without occupying precious context space for every single capability. Full details are only loaded when truly needed, striking a balance between capability breadth and context efficiency.

The Foundational Tools Skills Depend On

For skills to function properly, the agent needs two categories of foundational tools:

• File system access: The ability to read and write files;
• Code execution tools: The ability to run the various commands a skill requires in bulk.

Think of it this way: Skills provide the "knowledge and process," while these foundational tools provide the "hands and feet" for execution. Only together can they truly complete a task.

Flexible Combinations with MCP and Subagents

Skills don't operate in isolation — they can be combined with MCP (Model Context Protocol) and Subagents to build modular agent workflows.

MCP Background: MCP (Model Context Protocol) is an open-source communication standard released by Anthropic in November 2024, designed to address the fragmented integration between AI models and external data sources and tools. Before MCP, every new data source or tool (such as a database, API, or file system) required developers to write custom integration code, and the M×N combinatorial explosion made maintenance costs extremely high. MCP uses a client-server architecture that defines unified interfaces for resource exposure, tool invocation, and prompt injection, reducing complexity from M×N to M+N. Multiple development tools including Cursor, Zed, and Replit have already announced MCP support.

A Typical Combination Pattern

A concrete scenario looks like this:

• The agent uses MCP to fetch data from an external data source;
• Relies on a Skill to determine how to process or efficiently retrieve that data;
• Delegates specific tasks to a Subagent with an isolated context, where that subagent can itself invoke skills to access specialized knowledge.

Subagents and Isolated Contexts: The Subagent architecture is a core pattern in current multi-agent system (MAS) design. Its core value lies in context isolation — each subagent has its own independent context window, is responsible for a specific subtask, and returns results to the orchestrator upon completion. This design addresses the problems of context pollution and length overflow that a single agent faces in long-running tasks, while also allowing multiple subtasks to be executed in parallel for greater efficiency. A subagent's failure also doesn't contaminate the main agent's working memory, improving the overall system's robustness. This pattern closely resembles microservices architecture in software engineering, reflecting the evolution of agent engineering from a "monolithic" to a "distributed" architecture.

"MCP fetches the data, Skills understands the process, Subagents handle isolated execution" — this division of responsibility forms a modular, reusable agent architecture and is one of the most important trends in agent development today.

The Course's Hands-On Path: From API to SDK

This course emphasizes hands-on practice and covers the complete development chain from API to SDK:

1. Starting with the Claude API: Create a skill for a marketing campaign and combine it with pre-built Excel and PowerPoint skills;
2. Content creation and data analysis: Build two independent skills for content creation and data analysis workflows, and test them in the Claude API;
3. Code review and testing: Apply skills to Claude Code to implement automated code review and testing;
4. Building a research agent: Use the Claude Agent SDK to build a research agent that integrates and outputs research findings through skills.

About the Claude Agent SDK: The Claude Agent SDK is a high-level abstraction toolkit Anthropic provides for developers, encapsulating underlying details such as the agent loop, tool invocation, and multi-turn conversation management, so developers can focus on business logic rather than infrastructure. Compared to calling the raw API directly, the SDK provides more structured agent orchestration capabilities, supporting advanced features like tool registration, task decomposition, and result aggregation. The course's design — starting from raw API calls and gradually transitioning to the SDK — reflects the engineering education philosophy of "understand the underlying mechanics first, then use higher-level abstractions." This helps developers deeply investigate issues when they arise, rather than being completely in the dark about framework behavior.

This learning path spans multiple scenarios including programming, research, and data analysis, allowing learners to deeply understand the practical value of skills within real-world workflows.

Summary

The launch of Agent Skills is fundamentally about solving a recurring core pain point in agent development: how to enable agents to reliably and repeatedly execute specialized workflows. By encapsulating knowledge and processes into standardized skill files, developers can "build once, deploy anywhere."

Combined with the progressive disclosure mechanism for context management and flexible integration with MCP and subagents, Skills is becoming an indispensable infrastructure layer in the agent ecosystem. For engineers looking to dive deeper into agent development, understanding and mastering this system is a critical foundation for the next stage of development practice.

Key Takeaways