Scott and Mark learn to... live again | BRK543
Mark Russinovich and Scott Hanselman use live demonstrations and candid discussion to examine how developers can retain control as AI takes on more software development work. The session connects rapid prototyping with practical questions about judgment, review, operating costs, and learning.
AI-assisted development and developer judgment
The discussion opens with the pace of AI-driven change in software engineering, then considers the tension between excitement about coding tools and concerns about production quality. Frontier models, private local inference, and agentic coding frame a broader question: where intelligence runs, and how much control developers retain.
The session examines what AI can build today and where autonomous systems still fail. Its focus extends beyond generating code quickly to understanding and assessing the resulting software.
Decision models and reinforcement learning
The presenters introduce decision models and “system-one” AI thinking in the context of efficient agentic workflows. An embedding-based decision-model segment is paired with a reinforcement learning demonstration using Doom gameplay.
These demonstrations bring model-driven decision-making into the discussion alongside coding agents, rather than treating AI-assisted development solely as code generation.
Rapid prototyping and production risks
Agentic coding demonstrations cover quickly building AI-powered projects, including haptics experiments. The discussion then turns to the risks of shipping vibe-coded applications at scale, with lessons from Deskland and Semantic Start.
The distinction between a working prototype and production-quality software is a recurring concern: speed of construction does not resolve questions about quality or developer judgment.
Tools for inspecting AI-assisted work
Three demonstrations focus on understanding the systems and code involved:
- GPU Task Manager: A demonstration accompanied by discussion of platform APIs, heuristics, and extensibility.
- Agent Session Insights: A tool for analyzing coding sessions, developer judgment, and AI costs.
- Agent Review: An interface for understanding AI-generated code changes and speeding up consensus during review.
Together, these segments shift attention from what an agent produces to how developers inspect its work and assess the cost of a coding session.
Learning and the junior developer pipeline
The closing discussion addresses concerns about junior developer pipelines, mentorship programs, and the future of AI-assisted learning. It places the development of human judgment alongside the growing capabilities of coding agents.
Session chapters
- 00:00: Introduction and rapid AI-driven changes in software engineering
- 02:10: Balancing excitement and fear around AI coding, judgment, and production quality
- 05:10: Decision models, system-one AI thinking, and efficient agentic workflows
- 07:11: Embedding-based decision models and reinforcement learning with Doom
- 10:32: Building AI-powered projects quickly and experimenting with haptics
- 15:04: Risks of shipping vibe-coded apps at scale; lessons from Deskland and Semantic Start
- 21:21: GPU Task Manager, platform APIs, heuristics, and extensibility
- 27:05: Agent Session Insights: coding-session analysis, judgment, and AI costs
- 33:10: Agent Review: understanding generated code changes and speeding up consensus
- 39:11: Junior developer pipelines, mentorship, and AI-assisted learning
Session information and resource
This is Microsoft Ignite 2026 session BRK543, an intermediate (200) breakout in the Agent Platform track, presented in English (US).