Artificial intelligence is changing software development at a speed we have rarely seen before.Developers have already moved from writing every line of code themselves to working with AI assistants that can generate code, explain unfamiliar systems, create tests, debug applications, and automate repetitive work.But according to Microsoft MVP Andre Baltieri, that is only the beginning.In this episode of M365.FM, Mirko Peters sits down with Andre for a deep dive into the transition from traditional software development to AI-assisted development, coding agents, agentic architectures, Microsoft Agent Framework, .NET, RAG, context engineering, security, and the economics of generative AI.
FROM .NET IN 2003 TO THE AI ERA
Andre takes us back to the early days of .NET and C#, when learning a new Microsoft technology often meant purchasing official training and traveling to another city.Since then, software development has moved through desktop, web, mobile, cloud, containers, microservices, and serverless computing.Andre argues that the transition to AI feels fundamentally different. Instead of simply introducing another platform or framework, AI introduces a new way for humans to interact with software.
AI ASSISTANTS VS. AI CODING AGENTS
There is an important difference between having an AI assistant inside your IDE and delegating work to an agent.An assistant can explain code, suggest refactoring, answer questions, and help developers understand their applications.An agent can receive a goal, create a plan, divide the work into smaller tasks, use tools, coordinate additional agents, and implement significant parts of the solution.Andre explains how this is already changing his own development workflow, with AI now generating much of the code he previously would have written manually.
SPEC-DRIVEN SOFTWARE DEVELOPMENT
As agents become more capable, specifications become increasingly important.Instead of describing every implementation detail, developers can define requirements, architecture, constraints, and expected behavior and allow agents to determine how parts of the implementation should be completed.This shifts developer attention from simply producing code toward defining what should be built and why.
MICROSOFT AGENT FRAMEWORK
The conversation moves into Microsoft Agent Framework and its role in bringing AI capabilities into existing applications.Andre explains how the framework brings together capabilities associated with Semantic Kernel and AutoGen and provides developers with tools for connecting models, orchestrating workflows, using MCP, implementing RAG, handling data ingestion, and exposing application functionality to AI.For .NET developers in particular, this can significantly reduce the amount of integration code required.
WHY .NET STILL MATTERS IN THE AI ERA
Python remains one of the dominant languages in AI development, but Andre argues strongly that .NET and C# are extremely well positioned for enterprise AI applications..NET continues to evolve rapidly, while Microsoft's AI tooling increasingly gives C# developers native access to modern AI capabilities.Organizations with years of business logic already implemented in .NET may therefore have a major advantage: they do not necessarily need to rebuild everything before introducing AI.Existing functionality can instead be selectively exposed to agents and AI-powered applications.
FROM DETERMINISTIC SOFTWARE TO AGENTIC SYSTEMS
Traditional applications are largely deterministic:If X happens, execute Y.Agentic systems introduce another model:Here is the goal. Determine which actions are required to accomplish it.That represents a significant architectural shift.Instead of explicitly defining every possible path, developers increasingly define goals, tools, context, permissions, constraints, and boundaries within which AI can operate.
DESIGN PATTERNS ARE NOT DEAD
AI-generated code does not eliminate decades of software engineering knowledge.Clean code, maintainability, testing, architecture, and design patterns remain important because AI frequently learns how to implement new functionality by examining the existing codebase.Messy code can therefore lead to more messy code.Developers still need to understand architecture and engineering principles even when an AI agent performs much of the implementation.
THE STOCHASTIC SOFTWARE PROBLEM
Traditional developers expect identical inputs to produce identical outputs.Generative AI is probabilistic.The same request can produce different implementations, answers, or behavior across multiple executions.Andre discusses why this requires developers to rethink testing and validation and why strong guardrails become increasingly important when AI functionality is exposed to large numbers of users.
CONTEXT ENGINEERING IS MORE IMPORTANT THAN PROMPTING
Choosing the latest model is not necessarily the most important decision.Andre argues that context is everything.Developers need to understand both the business problem and the technical environment well enough to provide AI with the right information.Too little context produces weak results.Too much context can overwhelm the model.The challenge is finding the information that actually matters.
RAG, DATA AND THE CONTEXT WINDOW
Retrieval-Augmented Generation becomes especially important when organizations want AI systems to work with their own knowledge.But building a RAG system is not simply about putting documents into a vector database.Data needs to be cleaned, structured, chunked, retrieved, and inserted into the model's context intelligently.Andre shares an example from his own education platform, where video lessons were transcribed and indexed so users could search for concepts and jump directly to the relevant point in a video.
MEMORY AND MANAGING AI CONTEXT
Long-running AI conversations create another challenge: memory.As context windows fill, conversations need to be summarized or compacted.Andre explains why developers should actively manage this process instead of assuming that an AI system will always preserve the most important information.Sometimes the best solution is surprisingly simple: finish a task, close the conversation, and start again with a clean context.Specifications and Markdown files can also provide persistent project context for coding agents.
SECURITY, PERMISSIONS AND LEAST PRIVILEGE
Giving an AI agent access to tools and company data creates significant security implications.Andre recommends treating agents according to principles similar to human identities: close everything by default and expose only what the agent genuinely requires.Instead of giving an AI system unrestricted database access, developers should expose carefully controlled functions that return only the information required for a particular task.This becomes particularly important when agents can read or modify enterprise data.
PROMPT INJECTION AND AI GUARDRAILS
Prompt injection creates a new attack surface for AI-powered applications.Users can intentionally manipulate prompts, attempt to retrieve information outside the intended context, consume company resources, or persuade an AI system to perform actions its designers never anticipated.The discussion explores the importance of system instructions, application-level restrictions, controlled functions, identity, permissions, and platforms such as Azure AI Foundry for establishing additional security boundaries.
AI FINOPS — DON'T USE GENERATIVE AI FOR EVERYTHING
One of the most practical lessons from the conversation is that just because AI can perform a task does not mean AI should perform that task.Andre distinguishes between generative and deterministic workloads.If something must happen the same way every time, traditional programming may be faster, cheaper, and more reliable.He gives the example of his video workflow: Python scripts can extract audio and perform deterministic processing locally, while generative AI is reserved for tasks such as translation where generation actually adds value.The result is a hybrid architecture that can dramatically reduce unnecessary token consumption.
BUILDING THE AGENTIC FUTURE
Software development is moving beyond developers manually defining every individual step.Increasingly, developers will define goals, specifications, context, tools, permissions, architecture, and guardrails while AI systems determine how portions of the work should be accomplished.That does not eliminate the developer.It changes where the developer creates value.Understanding the business, designing maintainable systems, controlling context, securing tools and data, validating AI-generated work, and deciding when not to use AI may become some of the most important software engineering skills of the agentic era.RAPID FIRESingle agent or multi-agent?For complex workloads, Andre sees significant potential in multi-agent architectures and sub-agents.Prompt engineering or context engineering?Context engineering.And what comes next?More capable models, more powerful agents, better code generation, stronger architectures, and continued evolution of the tools developers use to build software.We are still at the beginning of the generative AI era.
ABOUT THE GUEST
Andre Baltieri is a Microsoft MVP and software development specialist with more than two decades of experience in the industry.His work focuses on .NET, C#, artificial intelligence, Microsoft Agent Framework, software architecture, and modern AI-assisted development.In this conversation, he brings together more than twenty years of software engineering experience with a practical view of how AI agents are changing the developer profession.
Become a supporter of this podcast: https://www.spreaker.com/podcast/m365-fm-a-microsoft-mvp-podcast-by-mirko-peters--6704921/support.
