Microsoft Reactor Sydney

AI coding agents can now write substantial amounts of open-source software, but generated code still needs more than a plausible implementation and a passing test suite.

In this session, Carl will demonstrate a workflow in which AI writes code, creates a formal specification, and produces a machine-checked proof of correctness. He is applying this approach to RangeSetBlaze, an open-source Rust library, using Lean to state and prove key correctness properties.

The broader idea is language-independent: AI may make formal verification practical for ordinary software development by taking on much of the proof engineering itself.

The session will explore what this workflow looks like in practice, where it works, where it breaks down, and how formal proofs can complement testing and code review as AI takes on more of the coding.

Design smarter AI workflows with visual orchestration, token optimization and extensible agent capabilities.

Session Introduction
AI agents are working across larger codebases and increasingly complex systems, making visual orchestration and efficient context management essential for modern development.

In this session, you’ll explore how to design smarter, token-optimized AI workflows using Canvases in the GitHub Copilot app, context engineering, and extensible agent capabilities.

Through practical demonstrations, you’ll see how developers can coordinate multi-agent work, manage context more effectively, and move beyond linear chat prompts to visual agent orchestration.

You Will Learn
• How to use Canvases in the GitHub Copilot app as a visual workspace for complex, multi-agent development
• How token budgeting, prompt caching, and context trimming can improve workflow efficiency
• How to extend agent workflows with Agent Plugins and Model Context Protocol (MCP)
• Best practices for moving from one-shot prompting to collaborative agent orchestration

Technologies Used
• GitHub Copilot harnesses including CLI, app, and Visual Studio Code
• Agent Plugins and Model Context Protocol (MCP)
• Token optimization and context engineering

Who Should Attend
• Developers building agentic workflows across complex codebases
• Engineers looking to improve AI workflow efficiency and context management
• Teams extending GitHub Copilot with reusable agent capabilities
• Anyone interested in visual, multi-agent development experiences

📌 This session is part of a series. Explore the full series here: https://aka.ms/AIGeniusS6

Measure, benchmark, and continuously improve Microsoft Foundry AI Agents using Microsoft IQ, Evaluations, and the Agent optimiser.

Session Introduction

Building an AI agent is just the beginning. Getting it right up front means giving it excellent context - but once it's live, keeping it good is a job of its own: evaluations to monitor quality, optimisation to keep cost in check, and reliable ways to tell whether any change actually helps

In this session, you'll learn how to measure, benchmark, and continuously improve AI workflow quality using Microsoft Foundry, Evaluations, and the Agent Optimizer, and how Microsoft IQ gives your agent excellent context. You'll see how continuous evaluation replaces guesswork with measurable signals across your agent's lifecycle.

You Will Learn
• How to compare and select models with Microsoft Foundry
• How to add organisational context to your Agents
• How to build continuous evaluation pipelines for your Agents
• How to profile the balance between quality, speed, and token cost
• How to use Agent Optimizer to improve your Agents
• How to manage your Agent in a CI/CD pipeline

Technologies Used
• Microsoft Foundry
• Microsoft Foundry Extension for Visual Studio Code
• Microsoft Foundry Evaluations and the Agent Optimizer
• GitHub actions to integrate with Foundry

Who Should Attend
• Anyone interested in building, evaluating and optimizing Agents, in Microsoft Foundry
• Teams introducing Agents into GitHub and CI/CD processes

📌 This session is part of a series. Explore the full series here: https://aka.ms/AIGeniusS6