Spec-Driven Development Over Vibe Coding

▶ Watch (02:00)

Spec Kit replaces one-line prompts with a Constitution file that sets governing principles (preferred cloud, MCP tools, frameworks) before any code runs. In the Spec Kit demo: specify, plan, task, and GitHub Issues (02:42), Shayne Boyer ran specify on a pet-listing app prompt. Spec Kit expanded that into user stories, an architecture plan, and a research file recommending GPT-4o for the chat feature.

The plan became individual GitHub Issues pushed via the GitHub MCP. Boyer assigned one to the Copilot Coding agent while handling front-end work himself. Spec Kit supports Claude Code and Gemini CLI, not only GitHub Copilot.

Building Agents with the AI Toolkit and Agent Framework

▶ Watch (10:52)

Rong Lu opened the AI Toolkit model catalog inside VS Code and asked Copilot which Azure AI Foundry model fits a pet-sitter recommendation agent. Copilot called the toolkit tool, returned cost, context window, and use-case details for each option, and recommended GPT-4o. One follow-up prompt generated a full working agent with system instructions, tool definitions, and tracing enabled via the Microsoft Agent Framework.

“Because models are the foundation for powerful agents.” — Rong Lu

Lu then prompted Copilot to orchestrate a second agent for pet venues in the multi-agent workflow demo: sitter and venue agents (14:28), using the framework’s workflow builder to route queries to the right agent.

Evaluating Agent Quality Before It Reaches Production

▶ Watch (15:53)

Agent outputs are non-deterministic, so testing them requires a different approach.

“So evaluations are kind of like unit tests” — Rong Lu

Lu sent a single prompt, “add evaluation to my agent,” and Copilot analyzed the workspace, suggested relevant metrics, and generated a test dataset visible in Data Wrangler. The evaluation script runs via the Azure AI SDK, either locally during development or as a GitHub Actions step triggered on every commit. That CI hook catches quality regressions before any deployment. Both single-agent and multi-agent comparisons are supported, and the GitHub Action installs directly from the GitHub Marketplace.

Automated Incident Response with the Azure SRE Agent

▶ Watch (21:44)

The Azure SRE agent runs a runbook as code. Shayne Boyer configured an execution plan: check CPU and memory, scan Bicep configurations for drift, attempt automated scaling, and open a GitHub Issue when human review is needed. The agent reads Azure resource groups and the app’s code repository before investigating anything.

In the SRE agent incident triage and pull request (24:51), the agent detected out-of-memory exceptions and replica scaling set to 1 instead of 3-5. It opened a pull request fixing both the config and memory allocation, used Playwright for before-and-after snapshots, and back-channeled the report through GitHub MCP. No engineer was paged.

Notable Quotes

Because models are the foundation for powerful agents. Rong Lu · ▶ 10:57

So evaluations are kind of like unit tests Rong Lu · ▶ 15:53

What we just showed you. This is all real. This is today. Amanda Silver · ▶ 30:44

Key Takeaways

  • Spec Kit converts a single natural-language prompt into user stories, a plan, and actionable GitHub Issues assigned to Copilot agents.
  • The Microsoft Agent Framework lets you build single agents or multi-agent workflows in VS Code, then evaluate them on every CI commit.
  • The Azure SRE agent reads a runbook, detects live incidents, and opens a pull request with config and code fixes — no on-call page required.