Agentic DevOps

From Editor to Production with GitHub Copilot

Microsoft Reactor • 60 Minutes

Why Agentic Changes Everything

DevOps was never about YAML files and pipelines.

Optimize the whole line, not just one station.

The Bottleneck Shift

Before

Code Gen
→
Review
⚠
Security
⚠
Deploy

Fast code generation. Slow everything else.

After

Code Gen
→
Review
→
Security
→
Deploy
✦ One Agentic Layer

The agentic layer smooths the whole flow.

The Contributor Graph

Top contributors to github/github

#1
🤖 Copilot Coding Agent
#2
👤 A Human
#3
🔍 Copilot Code Review

That's production. Not a demo.

Three Paths to Create

🌐

Coding Agent

Assign an issue, walk away.

💻

Agent Mode

Hands-on in VS Code.

⌨️

Copilot CLI

Terminal-native agent.

Pair vs. Peer Programming

Pair Peer
Prompt "Give me a function that sorts this array" "@test-agent, these three endpoints need tests. Use patterns in /tests/api/"
Style Micromanaging Delegating

The jump is huge: from micromanaging to delegating.

The 6 Elements of an Effective Agent

🎭 1

Specific Persona

⚡ 2

Executable Commands

📋 3

Example Output

🚧 4

Always / Ask / Never

🔧 5

Tech Stack + Versions

🧠 6

Context & Knowledge

This is your framework. Make it memorable.

Look at That

While we worked, Coding Agent finished.

✓ Read the issue
→
✓ Explored the repo
→
✓ Found state
→
✓ Created undo logic
→
✓ Updated UI
→
✓ Wrote summary

Async handoff. Assign, work elsewhere, return to draft PR.

Code Review

Copilot Code Review is already analyzing — custom instructions + CodeQL under the hood.
7x

Faster XSS fixes

12x

Faster SQL injection fixes

Custom Agents

Repetitive tasks? Automate them.

♿ Accessibility audits
📝 Documentation
🧪 Test coverage

One markdown file per agent.

No infrastructure. No deployment.

What Makes an Agent Work?

Most agents.md files? They don't work.

"You are a helpful coding assistant. Follow best practices."

Like telling a junior "write good code."

Agents need examples, limits, executable commands.

1. Persona & 2. Commands

Specific Persona

Not "helpful assistant." → "You are a test engineer specializing in React and Node.js applications."

Role clarity equals better decisions.

Executable Commands

Short commands at the top of the file:

@test component [path] @test api [endpoint] @test coverage
Commands over paragraphs. Always.

3. Expected Output

Don't explain how you want the code. Show it.

Arrange — prepare the user data

Act — render the component

Assert — verify the text appears

One example is worth more than three paragraphs of instructions.

4. Always / Ask / Never

The most important element. Three-tier boundaries.

✅ Always

Do without asking

Use Jest, follow AAA, aim for 80% coverage

🤔 Ask

Check first

Before testing third-parties, before modifying shared fixtures

🚫 Never

Hard limits

Modify source code. Delete failing tests. Commit to main.

The agent interpreted 'fix' as 'make them pass.' It deleted the failing tests.

Boundaries aren't bureaucracy. They're scars turned into rules.

5. Tech Stack & 6. Context

Tech Stack with Versions

Not "we use React." → React 18.2.0, Jest 29.5.0, TypeScript 5.1.0

Specific versions prevent outdated suggestions.

Context and Knowledge

Architecture: "Monorepo with Turborepo, /packages/ui for React components..."

✅ Can write: /tests 👀 Can read: anywhere 🚫 Cannot touch: source code
Understanding architecture = 70% fewer integration errors

The Agent Team

📄
@docs-agent

Documentation

Never modifies source code
🧪
@test-agent

Testing

Never deletes failing tests
🔌
@api-agent

Endpoints

Asks before schema changes
🚀
@dev-deploy-agent

Deploys to dev

Never deploys to production

Specialists with boundaries.

Agent Chaining

💻

Develops

→
🧪

Tests

→
📝

Documents

→
🔍

Reviews

Assign issue. Go to lunch. Return to draft PR with tests and docs.

The SDK

IssueCrush is built on the Copilot SDK.

🔎 Analyzes issues
📋 Generates summaries
🔗 MCP integration

BYOS: Bring Your Own Subscription

Uses your Copilot subscription. Open source.

Your subscription isn't autocomplete. It's infrastructure.

Agent HQ Announcements

Pick your agent. All available today.

🤖

Copilot

GitHub native

🧠

Claude

Anthropic

⚡

Codex

OpenAI

🔮

More coming

Stay tuned

The Reality

70% AI
30% Human

🤖 AI Territory

  • Basic CRUD
  • Boilerplate tests
  • API documentation
  • Tech debt cleanup

Incredibly fast

🧑‍💻 Human Territory

  • Legacy system integration
  • Edge cases
  • Debugging staging issues
  • Architecture decisions

Requires judgment

"AI gives you speed. You bring the craft."

Where This Is Going

We used Copilot to build Copilot.

March sprint: Coding Agent was one of the most productive contributors to Coding Agent itself.

🧹 Tech debt
🧪 Test coverage
📝 Documentation

Humans focus on architecture and decisions.

What We Covered Today

🛠️

Create

Coding Agent, Agent Mode, CLI

🤝

Collaborate

Code Review, Custom Agents, The 6 Elements

⚙️

Operate

SDK, BYOS, Production Reality

One agentic layer across the whole lifecycle.

Three Things to Remember

🏗️

Fundamentals First, Always

Clear persona, examples, explicit boundaries. Not magic prompts copied from Twitter.

📊

Measure Experience, Not Just Output

High velocity without quality isn't productivity. It's technical debt with better marketing.

🤝

Fluency Over Dependency

The best teams know when NOT to use agents. That's the real skill.

Resources

The agents that work aren't magic.

They're specialists with clear instructions.

Fundamentals first, always.

Questions?