Why your AI code doesn’t ship: Closing the gap to production | BRK200

Name: Why your AI code doesn’t ship: Closing the gap to production | BRK200
Uploaded: 2026-06-03T10:02:13+00:00
Description: Mario Rodriguez and Evan Boyle demonstrate how AI agents—centered around GitHub Copilot—can work across planning, coding, CI/CD, and production operations...

Today by Mario Rodriguez, Evan Boyle

Mario Rodriguez and Evan Boyle demonstrate how AI agents—centered around GitHub Copilot—can work across planning, coding, CI/CD, and production operations to help teams ship AI-assisted code safely. The session focuses on practical workflows, guardrails, and review patterns that keep autonomous changes controlled while still moving fast.

Overview

In this Microsoft Build 2026 breakout (BRK200), the speakers walk through a demo-heavy, end-to-end view of “agent-native” engineering: AI systems that can plan work, implement changes, open pull requests, interact with CI/CD, and assist with live operational tasks.

Key themes include:

Using AI agents beyond code completion (planning, PR creation, pipeline fixes, and production patching)
Keeping agents constrained with explicit workflows and review steps (“keep agents on a leash”)
Treating the dev lifecycle as a connected system where agents can collaborate across stages

Session structure (chapters)

Welcome and introduction

Context for Build 2026 and the shift toward AI agents operating across the software lifecycle.

Agents enter the workforce: a new collaboration model

Positions agents as collaborators that can take on tasks spanning multiple phases of delivery, not just IDE assistance.

Pillars for an agent-native GitHub engineering system

Introduces a set of guiding principles/pillars for building an engineering system where agents can contribute safely.

Demo: GitHub Copilot app and agent collaboration

Demonstrates workflows where GitHub Copilot is used as the center of multi-step work.

Copilot app workflows: creating and managing multiple sessions

Shows how work can be split into multiple sessions to manage context and parallelize tasks.

Canvases and custom workflows

Introduces “canvases” and how they can be used to structure work and define repeatable flows.

Whiteboard canvas creation and extension marketplace

Demonstrates creating a whiteboard-style canvas and connecting it to an extension marketplace model for expanding workflows.

Copilot cloud sandboxes and Chronicle for work continuity

Covers cloud sandbox environments for agent work and a “Chronicle” concept for continuity across sessions.

Agent merge: autonomous code reviews and integration

Demonstrates an “agent merge” flow where agents participate in code review and integration steps.

Automation, rubber duck, and multi-model security review

Shows additional patterns for agent-assisted automation, “rubber duck” style reasoning, and using multiple models for security review.

Wrap-up

Summarizes the agent-native engineering system approach and closes the session.