Introduction to Agent-First Development | Ep 1 of 6 - VS Code Learn

Overview

This is a tutorial-style introduction to getting started with “agent-first” development in VS Code using GitHub Copilot (Copilot Chat). The video explains core concepts and walks through a live session setup and configuration, showing how to turn an LLM into an actionable agent inside your editor.

Five-part formula of success: Harness, Model, Context, Tools, Prompts

Key concepts

Harness

• VS Code + GitHub Copilot Chat acts as the harness that turns an LLM into an agent.
• The harness exposes model selection, agent modes, enabled tools, and access to files.
• Analogy: like a wiring harness in a car — it routes model power to the right components.

Model

• Multiple model choices are available (examples mentioned: Sonnet, Codex).
• “Thinking effort” settings (low / medium / high) trade off speed vs depth of reasoning.
  • Use high for architecture, debugging, or complex reasoning.
  • Use low for simple boilerplate or fast iterations.

Context

• Agents need explicit context from your repo to produce code specific to your codebase.
• VS Code Copilot Chat can automatically read directories or you can manually attach context (files/folders, GitHub resources, MCP resources, etc.).
• You can reference files by name (for example, using a #filename convention).

Tools

• A tool picker lets you enable/disable built-in and extension-provided tools:
  • Examples: to-dos, terminal/command execution, file edit/read, search, integrated browser, run commands, delegate to other agents.
• Configure tools granularly — only enable what’s necessary for the session.
• Agent actions (creating todos, running commands) depend on the tools you enable.

Prompts

• Craft prompts with enough specificity to avoid vagueness without over-constraining implementation details.
• Prompt quality is crucial for getting useful agent output.

Product features & UI walkthrough

• VS Code layout customization: move primary sidebar left/right and change activity bar position.
• Demonstration steps shown:
  • Selecting a model (e.g., Sonnet).
  • Setting thinking effort to high.
  • Pasting a prompt and starting an agent session.
• Tool picker demo:
  • Pruning a large set of tools/extensions down to a focused set.
  • Enabled tools produce matching UI icons (e.g., todo list, terminal).
• Permissions:
  • Agents may request permission to run terminal commands.
  • Approval prompts are shown; details on approval levels are promised in the next video.

Practical guidance / checklist

To start an agent-first session, do the following:

• Customize your VS Code layout as desired.
• Choose the right model and thinking-effort level for the task.
• Craft clear, targeted prompts.
• Configure tools — enable only those the session needs.
• Attach relevant files/folders or other context sources to the session.
• Monitor and approve agent actions (permissions).

What to expect next

The next episode will cover permission and approval levels for commands and show how to allow or skip specific agent-run actions.

Speakers / sources

• VS Code Learn instructor / presenter (unnamed)
• Demonstrated products/services:
  • Visual Studio Code
  • GitHub Copilot (Copilot Chat)
  • Sonnet model
  • Codex (mentioned)
  • MCP service/extensions