Build apps & agents that scale with VS Code, GitHub Copilot, and Agent Framework

Quick Overview

The presentation highlights the evolution of agentic workflows, exemplified by building and deploying a multi-agent system for pet-sitter recommendations using GitHub Copilot and Azure AI services, culminating in the announcement of the Azure SRE Agent for continuous production improvement.

Key Points: GitHub Copilot was used to build a pet-sitter recommendation application from scratch, demonstrating agentic workflows from specification to deployment. The presentation introduced the concept of evolving developers into AI orchestrators who manage teams of specialized AI agents. GitHub Spec Kit was shown as a tool to create unambiguous, specification-driven development, contrasting with traditional, manually updated, infrastructure-heavy code. The Azure AI Agent Evaluation tool was demonstrated to provide offline evaluation of AI models and agents within CI/CD pipelines, showing metrics like Tool Call Accuracy, Response Relevance, and Response Coherence. The final stage involved deploying a custom Azure SRE Agent, which automates incident management and resource optimization based on predefined execution plans and tools. A key metric shown was the 80.0% Tool Call Accuracy for the Sitter Agent evaluation run.

Context: This presentation segment, delivered during GitHub Universe '25, focuses on the maturation of AI-assisted development, specifically showcasing how GitHub Copilot and the Microsoft Agent Framework facilitate the creation, evaluation, and deployment of complex, multi-agent systems. The core theme is the shift from developers writing infrastructure-heavy code to developers specifying behavior and orchestrating AI agents to execute tasks, exemplified by building a functional pet-sitter recommendation app.

Detailed Analysis

The presentation details the progression toward AI-driven development, moving from agents that rely on fixed, manually updated logic to autonomous, specification-driven systems where developers act as orchestrators collaborating with AI agents. GitHub Spec Kit is presented as the tool enabling this shift by allowing developers to define clear, testable specifications that guide agent behavior, rather than writing verbose, platform-specific code. The process demonstrated involved using Copilot to build the 'Octopets' application, which finds pet-friendly venues. This development workflow included utilizing Azure AI tools, specifically the Azure AI Agent Evaluation framework, to run offline testing via CI/CD pipelines. The evaluation results showed metrics like 80.0% Tool Call Accuracy and 100% Response Relevance/Coherence, confirming the agent's capability. Finally, the Azure SRE Agent was introduced as a tool to automate incident management and resource scaling based on these evaluations and predefined response plans, significantly reducing manual toil and improving system reliability. The presenter concluded by showing the successful deployment and evaluation of the SRE Agent itself.