Subagents, Agent Skills, and MCP

In this episode we’re going to go a little deeper with coding agents by learning how to customize their functionality through advanced features. Subagents, Agent Skills, and the Model Context Protocol (MCP) all represent ways to extend what a coding agent can do beyond its default configuration — delegating work to other agents, packaging reusable expertise, and connecting to external tools and data sources.

Before getting into these topics, let’s take a moment to better understand the anatomy of a coding agent session. Even before you enter your first prompt in a new OpenCode session, the session context already includes a lot of information, such as:

System prompt — instructions from the agent’s developers that define its role, behavior, and constraints. (e.g. “You are opencode, an interactive CLI agent specializing in software engineering tasks. Your primary goal is to …”).
Tool definitions — descriptions of the tools available to the agent (reading and editing files, running shell commands, fetching web pages, etc.) and guidance on when to use them.
Project context — files like AGENTS.md (see the previous episode) that are read automatically so the agent starts out already familiar with your project’s conventions and structure.

Subagents

A subagent is a separate agent instance created by the main AI coding agent to handle a specific task. Subagents have their own system prompts and they run in a separate context, not in the main conversation thread.

The subagent will start in a new session with its own context window, receive instructions from the main agent, complete the task, and return a summary to the session that spawned it. This is another advantage of the subagent: it won’t overload your current chat session with all of the intermediate analysis and tool calls; it only returns a summary once the task is completed.

Exploring System Prompts

OpenCode includes two pre-configured subagents called Explore and General. We can learn about these subagents by reading their system prompts.

Start a new session and try the following prompts in succession.

Print the first few sentences of your system prompt
@general Print the first few sentences of your system prompt
@explore Print the first few sentences of your system prompt

Creating a Subagent

OpenCode has a built-in tool to create subagents:

opencode agent create

Challenge: A dependency analysis subagent

Create a subagent to analyze a codebase and report dependencies.

Agent Skills

Agent Skills are a way to encapsulate and distribute specialized knowledge or workflows for agents. The core specification is described at https://agentskills.io.

A minimal Agent Skill consists of a directory with a file, SKILL.md, that includes some metadata (a name and description in the file’s front matter) and some content in the body of the file.

For example, here is a portion of the SKILL.md file from an Agent Skill describing how to make best use of the Coder Workspace environment. (The full text is available on GitHub).

---
name: dreamlab-coder-workspace
description: Use features of the DREAM Lab Coder Workspace Environment
---

You are running in an Ubuntu-based virtual machine with sudo access. 

[etc..]

Skills are meant to be shared between projects. Sites like https://www.skills.sh are useful for searching available skills. Be warned: skills have been used to distribute malware! It’s very important to review the source code for a skill before installing it.

Installing Skills

Like many coding agent configurations, you can install Agent Skills in either a project-specific or global configuration directory. (See OpenCode’s Agent Skills documentation).

We will use the following: - Project agent-compatible: .agents/skills/<name>/SKILL.md - Global agent-compatible: ~/.agents/skills/<name>/SKILL.md

You can install skills manually, but we’ll use Vercel’s command line utility called, appropriately enough, skills.

Note

The skills utility requires the NodeJS Package Manager, which may not be installed on your Coder Workspace. To install it:

sudo apt update -y && sudo apt install -y nodejs npm

Let’s install a few Agent Skills from the DREAM Lab (see the code: https://github.com/ucsb-dreamlab/skills):

npx skills add ucsb-dreamlab/skills

Use the text interface to install the dreamlab-coder-workspace and web-a11y skills to your local project. You will need to restart any OpenCode sessions to enable the skills. You should see the new skills listed when you use OpenCode’s /skills command.

Using Skills

When a skill is installed, only its name and description are loaded into session context. The LLM may choose to ‘invoke’ the skill, in which case the body of the SKILL.md file is loaded into context. This doesn’t always work, however! If the agent doesn’t load a relevant skill on its own, you can invoke the skill explicitly with the /skill-name command, for example:

/dreamlab-coder-workspace what is the public URL for my Shiny dashboard?

Model Context Protocol (MCP)

Model Context Protocol provides a way to give agents additional tools, beyond those that may be built-in. For example, you can augment OpenCode’s built-in tools with specialized capabilities.

A very common use of MCP is to give coding agents the ability to directly control a web browser.

The Coder Workspace skill we installed above includes instructions for installing a “headless” web browser (a browser that can be used on a computer without a monitor or a graphical interface) and enabling the chrome-devtools MCP server.

Try the following prompts:

/dreamlab-coder-workspace enable the chrome-devtools mcp server for opencode

When it’s done, restart OpenCode and use the /mcp command to confirm the chrome-devtools MCP server is installed. Then, try the following prompt:

Take a screenshot of my Shiny data dashboard and describe it to me