AI-assisted development is moving beyond autocomplete and chat. Modern AI coding agents can inspect repositories, modify files, run tests, analyze failures, create pull requests, and execute increasingly complex development tasks. In this episode of the M365 FM Podcast, Mirko Peters talks with Microsoft MVP Lars Gyrup Brink Nielsen about GitHub Copilot, AI coding agents, developer experience, automated testing, TypeScript, JavaScript, Nx, monorepos, Polygraph, GitHub Actions, Playwright, and the architectural foundations required to make agentic software engineering work at enterprise scale.
FROM AI ASSISTANCE TO AI CODING AGENTS
For the last few years, AI development tools have primarily helped developers generate code, autocomplete functions, answer technical questions, and accelerate individual programming tasks. Agentic engineering changes that model. AI coding agents can potentially inspect an existing repository, understand an issue, modify multiple files, run automated tests, analyze failures, make corrections, and prepare a pull request. That changes the central question from "Can AI generate code?" to "Can AI safely perform engineering work inside a real software architecture?"
MEET LARS GYRUP BRINK NIELSEN
Lars Gyrup Brink Nielsen is a Microsoft MVP in Developer Technologies, author, international speaker, tech writer, open-source maintainer, community organizer, and former GitHub Star. His experience spans frontend development, cloud-native systems, developer experience, automated testing, continuous delivery, deployment, and open-source software. He has also worked with GitHub Copilot since its early days and has spent more than a year experimenting with AI coding agents in open-source development.
WHY DEVELOPER EXPERIENCE MATTERS
Developer experience is fundamentally about removing friction from software development. That includes choosing and integrating development frameworks, testing tools, build systems, CI pipelines, documentation, reusable packages, development environments, architectural standards, and automation. Lars contrasts modern developer tooling with earlier workflows involving FTP deployments, directly editing production systems, shared ZIP files, and environments without proper source control. Modern developer experience replaces these fragile processes with repeatable, automated, and verifiable workflows.
ARCHITECTURE SHOULD BE ENFORCED
Documentation alone isn't enough for large software architectures. Lars explains how tooling such as Nx can attach metadata to projects and enforce architectural boundaries. Organizations can define which types of projects may depend on other projects and use linting rules to automatically detect violations. Generators can also scaffold components, services, projects, tests, and configurations according to organizational standards. This becomes particularly important with AI coding agents. Instead of expecting an agent to remember every architectural rule, organizations can make those rules automatically enforceable.
MONOREPOS EXPLAINED
A monorepo stores multiple projects or systems inside a shared source-control repository. Microservices and micro-frontends don't necessarily require separate repositories. They can exist within one monorepo while still being independently built and deployed. The advantage is that dependencies between projects become easier to understand and changes spanning multiple systems can be tested together. However, Lars also explains that moving hundreds or thousands of developers from established repositories into one monorepo can be extremely difficult organizationally.
MONOREPO VS MULTIPLE REPOSITORIES
Separate repositories don't eliminate dependencies between teams. They simply manage those dependencies differently. When two teams maintain dependent systems in different repositories, changes can require coordination, separate environments, cross-repository testing, and synchronized deployments. A monorepo can make those dependencies more visible because the relevant source code exists within the same repository. The larger challenge is often organizational rather than technical: teams still need to communicate, coordinate ownership, and manage dependencies regardless of repository strategy.
WHAT IS NX?
Nx is a development toolchain that helps teams manage complex codebases, project dependencies, tasks, testing, builds, linting, generators, and development workflows. Lars explains how Nx has evolved beyond its earlier JavaScript-focused roots. It can integrate development tools, automate migrations, understand project relationships, orchestrate tasks, and support multiple technologies. It can also determine which projects are actually affected by a change rather than unnecessarily rebuilding or retesting an entire large codebase.
CACHING CAN DRAMATICALLY ACCELERATE DEVELOPMENT
Large codebases can contain hundreds of projects and enormous test suites. Repeatedly running every build, test, linting operation, and compilation task wastes significant developer and CI time. Nx can cache task results. If the relevant source files and dependencies haven't changed, developers can reuse previous results instead of executing the same work again. Remote caching can also allow teams and CI systems to share those results. For AI agents, faster verification becomes especially valuable because an agent may repeatedly modify, test, inspect, and correct code during a single task.
POLYGRAPH AND MULTI-REPOSITORY AI AGENTS
Not every enterprise can move hundreds of existing repositories into a monorepo. Lars discusses Polygraph as a newer approach for helping AI coding agents operate across multiple related repositories. Instead of an AI coding session being isolated to one repository, Polygraph can provide a harness around multiple repositories and their dependencies. This can allow an agent to work across frontend, backend, microservice, and other repositories during the same development task and potentially prepare coordinated pull requests across those systems.
GITHUB ACTIONS AND CONTINUOUS INTEGRATION
Automation is a major part of the development environment discussed throughout the episode. GitHub Actions can provide the CI foundation for building, testing, validating, and deploying changes. Nx can work alongside CI workflows and delegate tasks through its own cloud capabilities. During the rapid-fire round, Lars gives GitHub Actions a particularly strong endorsement, calling it the best CI system.
PLAYWRIGHT FOR AUTOMATED TESTING
Automated verification becomes more important as AI writes a larger percentage of software. When asked to choose between Playwright and Cypress during the rapid-fire round, Lars chooses Playwright. The broader point is significant for agentic development: organizations need reliable automated feedback loops. If an agent changes code, it needs tools capable of determining whether those changes still satisfy the expected behavior.ㅤ
GITHUB COPILOT AND THE AGENTIC SHIFT
Lars was an early GitHub Copilot user through his involvement with the GitHub Stars program and provided feedback during Copilot's earlier development. His development workflow has now moved significantly beyond traditional Copilot autocomplete. He explains that he rarely writes code manually anymore and instead instructs coding agents to perform much of the implementation work. The challenge increasingly becomes reviewing and ensuring the quality of the code produced by those agents.
CONTEXT IS THE NEW ENGINEERING PROBLEM
An AI model may understand TypeScript, JavaScript, Angular, React, .NET, testing frameworks, and popular development tools from its training. What it doesn't automatically understand is your organization. Every large software environment contains internal naming conventions, architectural decisions, team standards, project-specific rules, historical constraints, and unique development practices. Those rules need to become accessible to AI agents through mechanisms such as agent instructions, skills, automated checks, architectural boundaries, and repository-specific context.
CLEAN CONTEXT FOR AI CODING AGENTS
Giving an AI agent more information isn't automatically better. Large repositories can contain enormous amounts of source code, documentation, configuration, legacy decisions, and irrelevant files. The challenge is providing the agent with the information required for its current task without overwhelming its working context. Development tooling, dependency graphs, repository structures, project metadata, agent instructions, and automated validation can help narrow that context. The objective is to give agents enough information to make architecturally correct decisions rather than merely locally correct code changes.
CODIFY YOUR TEAM STANDARDS FOR AI
Organizations don't necessarily need to fine-tune an AI model on their complete private codebase. Lars points instead toward codifying team-specific knowledge into agent skills and instruction files such as AGENTS.md. These instructions can explain how a particular team performs common tasks, which conventions should be followed, and how the codebase should be approached. This transforms knowledge that previously existed mainly inside experienced developers' heads into context AI coding agents can use. ㅤ
Become a supporter of this podcast: https://www.spreaker.com/podcast/m365-fm-modern-work-security-and-productivity-with-microsoft-365--6704921/support.
