Scott and Mark learn...how agents reshape software engineering | BRK247
Mark Russinovich and Scott Hanselman discuss how AI agents are changing day-to-day software engineering, focusing on where agentic workflows speed things up, where they break down, and what engineers can do to adapt without buying into hype.
Overview
This Microsoft Build 2026 breakout (BRK247) looks at how AI agents are reshaping software engineering practice. The session emphasizes practical realities: common failure modes, how engineers can interpret AI output safely, and how roles and skills may shift as agentic tooling becomes more common.
AI-augmented software practices (Project Lobster / Aspire team)
- The session opens with a demonstration framed around AI-augmented development practices.
- It references Project Lobster and the .NET Aspire team as part of the discussion and demo context.
AI compared to an intern: limits in context and learning
- The presenters compare AI behavior to an intern in terms of:
- Limited context about the full system.
- Limited ability to learn from experience in the way a human teammate does.
Failure modes: faulty fixes and benchmark misinterpretation
- The talk highlights examples where AI-generated fixes are incorrect or misleading.
- It calls out that benchmarks can be misunderstood or misapplied, leading to wrong conclusions about performance or correctness.
Demonstration-driven discussion: ZoomIt panorama feature
- The session uses a ZoomIt panorama feature scenario to illustrate challenges that can trip up AI-generated code and reasoning.
Why some problems are hard: ClearType and pixel color complexity
- The presenters discuss the complexity behind ClearType and pixel color challenges as an example of:
- Non-obvious constraints.
- Edge cases that require deep domain knowledge.
- Situations where “looks right” fixes can be technically wrong.
Pitfalls of AI-generated code and impact on early-career developers
- The session discusses how AI-generated code can introduce pitfalls, including:
- Incorrect assumptions.
- Superficially plausible changes that don’t match real requirements.
- It also touches on how these dynamics can affect early-career developers and learning pathways.
Historical perspective: technology waves and skills evolution
- The presenters frame AI as another technology wave that triggers anxiety, but also drives skill evolution over time.
Training analogy: guided real experiences and safe mistakes
- The session uses a “preceptor” analogy to describe how engineers can learn effectively:
- Through guided real-world practice.
- By making safe mistakes and learning from them.
Outlook: AI won’t replace oversight; focus shifts to learning and mentoring
- The presenters argue that human oversight remains necessary.
- They suggest the focus shifts toward:
- Learning and mentoring.
- Cognitive engagement and judgment.
- Adapting engineering practices to account for agentic tooling.