06.01 · GitHub Copilot & Coding Agents

Level: Intermediate
Pre-reading: 06 · AI Tool Ecosystem · 05 · MCP Servers

GitHub Copilot — Beyond Autocomplete

GitHub Copilot has evolved from line-by-line autocomplete into a full coding agent:

Mode What It Does
Code completion Next-line and next-block suggestions in the editor
Chat Ask questions about selected code, explain, refactor
Inline chat Edit code in-place with natural language instructions
Agent mode Multi-step tasks with MCP tool use, file creation, command running
Code review PR-level review and suggestion generation

Copilot Agent Mode Architecture

In agent mode, Copilot is a full ReAct agent with access to:

  • Filesystem tools — read, write, create files in the workspace
  • Terminal tools — run commands (build, test, lint)
  • MCP tools — any MCP server you configure in VS Code settings
  • GitHub tools — search repos, read PRs, check issues
graph LR
    A[Developer prompt in VS Code] --> B[Copilot Agent]
    B --> C[MCP: GitHub Server]
    B --> D[MCP: JIRA Server]
    B --> E[Filesystem tools]
    B --> F[Terminal: mvn test]
    C --> B
    D --> B
    E --> B
    F --> B
    B --> G[Code changes + PR opened]

Configuring MCP Servers in VS Code

// .vscode/mcp.json
{
  "servers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "${env:GITHUB_PAT}" }
    },
    "jira": {
      "command": "uvx",
      "args": ["mcp-jira"],
      "env": {
        "JIRA_URL": "${env:JIRA_URL}",
        "JIRA_TOKEN": "${env:JIRA_TOKEN}"
      }
    }
  }
}

Never commit tokens to git

Always use ${env:VAR_NAME} syntax which reads from your shell environment, not hardcoded values in the JSON file. Add .vscode/mcp.json to .gitignore if it may contain environment references tied to your user.

Coding Agent Patterns for Spring Boot

Task Agent Prompt Pattern
Implement endpoint "Implement a POST /orders endpoint in order-service that validates the request against OrderRequest DTO and persists via OrderRepository"
Fix failing test "Test OrderControllerTest.shouldCreateOrder is failing with NullPointerException. Find the root cause and fix it."
Refactor service "Refactor PaymentService to use the Strategy pattern for payment method handling. Keep existing tests passing."
Add observability "Add OpenTelemetry spans to all public methods in OrderService. Name spans with PRODUCT_NAME.SERVICE_METHOD format."

Aider — Git-Native CLI Agent

Aider is an open-source alternative that works entirely through git diffs:

graph LR
    A[aider: fix the NPE in OrderController] --> B[LLM generates unified diff]
    B --> C[aider applies diff to files]
    C --> D[git commit automatically created]
    D --> E[Developer reviews commit history]

Why it's interesting: Aider uses the git history as its state store — every change is a trackable commit. The developer can easily git revert if the agent went wrong. This aligns well with code review practices.

→ Deep Dive: Custom Agent Workflows — /analyze-requirement, /dev-agent, /review-pr

When should you use Copilot agent mode vs. a custom LangGraph agent?

Copilot agent mode is excellent for developer-interactive tasks in the IDE (implement a feature, fix a test, explain code). Custom LangGraph agents are better for fully automated pipelines triggered by CI/CD events, because you have full control over the loop, state, and human-in-the-loop gates. They're complementary: Copilot for developer workflow, LangGraph for automation pipelines.