I recently studied the AI Coder: Complete Claude Code & Coding Agents Course. Here are the main things I learned about AI coding agents, the tools they use, and how to make them work more effectively.
1. MCP (Model Context Protocol)
MCP is an open protocol that allows AI applications to connect to external tools and services.
For example, Claude Code can use MCP to connect to external services and get information or perform actions.
MCP is not a tool itself. It is a way for an AI agent to communicate with tools.
The main problem with MCP is context usage. When an MCP server provides many tools, their descriptions can take up a lot of the AI's context.
Three main MCP components
- MCP Host: The AI application, such as Claude Code.
- MCP Client: The part of the application that connects to an MCP server.
- MCP Server: A program that provides tools or information to the AI application.
2. Skills
Skills give AI agents extra knowledge and instructions for completing specific tasks.
For example, a skill can teach an agent how to review code, develop a feature, or follow a particular workflow.
Skills can save context because the agent does not need to load all the instructions at once. Instead, it can load more information when needed. This is called progressive disclosure.
Three levels of progressive disclosure
- Metadata: The skill's name and description.
- Instructions: Detailed steps explaining how to perform the task.
- Resources: Additional files and reference materials that the agent can access when needed.
Where to find skills
- Anthropic Skills: A collection of skills available on GitHub.
- skills.sh: A website for discovering and installing skills for AI coding agents.
The main limitation of skills: They provide instructions and knowledge, but they do not replace tools that perform actions or connect to external services.
3. Plugins
Plugins make it easier to add new capabilities to AI coding agents. A plugin can combine several things, such as MCP connections, skills, and commands.
Instead of setting everything up separately, you can install a plugin that provides the capabilities you need.
The downside is that plugins may use more context or offer less flexibility than configuring individual components yourself.
My recommendation: Start with plugins because they are easier to set up. If you need more control later, you can configure individual tools and skills.
Some plugins work only with specific AI coding agents, so always check compatibility before installing them.
4. Tools I Learned About
Context7
Context7 provides up-to-date documentation for programming libraries and frameworks. It helps AI agents find current documentation instead of relying only on what they learned during training.
Polygon.io
Polygon.io provides financial market data through APIs, including stock prices. It can be useful when building applications that need stock market information.
Feature-dev
Feature-dev provides a structured way to develop new features. Instead of immediately writing code, the agent follows steps to understand the codebase, plan the solution, implement it, and review the result.
Cerebras Inference
Cerebras provides infrastructure for running AI models. One of its key benefits is very fast text generation, which can make AI applications more responsive.
5. Subagents
Subagents are smaller AI agents that can work on specific tasks assigned by a main agent.
For example, one agent can investigate the code, another can implement a feature, and a third can review the changes.
Advantages:
- Parallel work: Multiple agents can work on different tasks at the same time.
- Separate context: Each agent can focus on its own task without putting everything into the main agent's context.
- Specialization: Each agent can have a specific responsibility.
Disadvantages:
- More complexity when coordinating agents.
- Higher token usage and potentially higher costs.
- Extra work to combine and review their results.
Subagents are most useful when a task can be divided into smaller, independent tasks.
6. Sandboxing
A sandbox is an isolated environment where an AI agent can run code and perform actions without having unrestricted access to the rest of the system.
Sandboxing improves security and can also make development faster. For example, if an agent is allowed to run certain commands safely, you do not need to approve every single action manually.
This helps avoid approval fatigue — when you become so used to approving requests that you stop checking them carefully.
Types of sandboxing
- Local sandbox: Runs on your computer, with restrictions on what the agent can access.
- Managed cloud environment: Runs in a cloud environment managed by a provider.
- Cloud sandbox: Runs in an isolated cloud environment that can support different coding agents.
The main idea is to give AI agents enough freedom to work efficiently while limiting the risks.
7. Spec-Driven Development (SDD)
Spec-Driven Development means writing down what you want to build before asking an AI agent to implement it.
For example, instead of asking an agent to build a booking system with a short prompt, you first describe its features, requirements, and expected behavior.
The agent then uses this specification to guide development.
This approach helps reduce misunderstandings and makes it easier to check whether the final result meets your requirements.
8. Get Shit Done (GSD)
GSD stands for Get Shit Done. It is a workflow designed to make AI-assisted development more structured.
The idea is to break a large project into smaller tasks, plan the work, and guide the AI agent through implementation.
This is useful when a project is too complex to complete reliably with a single prompt.
9. Agent Teams
Agent teams allow multiple AI agents to collaborate on a larger task.
For example, different agents can work on the backend, frontend, and testing.
The advantage is that work can be divided between agents, potentially saving time.
However, using more agents does not always produce better results. They need to be coordinated, and running multiple agents can increase costs.
Final Thoughts
Here are my main takeaways from the course:
- MCP connects AI applications to external tools and services.
- Skills give AI agents reusable instructions and knowledge.
- Plugins make it easier to add new capabilities.
- Subagents divide work between specialized agents.
- Sandboxing makes it safer for AI agents to execute code.
- SDD starts development with clear requirements.
- GSD breaks projects into smaller, manageable tasks.
- Agent teams allow multiple agents to collaborate.
- Cerebras can help make AI applications faster through low-latency inference.
My biggest takeaway is that AI-assisted coding is about more than choosing a good AI model. You also need to know how to provide the right instructions, tools, context, and environment.
These concepts help make AI coding agents more useful in real-world software development.