The Future of AI: Structured Vibe Coding - An Improved Approach to AI Software Development

Microsoft

Oct 13, 2025

The Future of AI blog series is an evolving collection of posts from the AI Futures team in collaboration with subject matter experts across Microsoft. In this series, we explore tools and technologies that will drive the next generation of AI. Explore more at: Collections | Microsoft Learn

One of the realities of my job is that I often face large, repetitive tasks - most notably, completing extensive questionnaires in sprawling Excel files. Each sheet often contains dozens, if not hundreds, of questions. Manually copying, pasting, and answering each question takes far too much time.

So, I set out to automate this part of my workflow.

Along the way, I discovered that the best way to build with AI agents is similar to the dynamics of managing a human engineering team: you start with a spec, assign issues, and let the team execute. We refer to this approach as structured vibe coding.

From Manual Work to Multi-Agent Systems

A while back, I built a small multi-agent system to help with questionnaire answering. It worked, but the process was clunky. I still had to manually copy questions into the tool and trigger agents by hand.

This time, I wanted a cleaner, end-to-end solution. So, I started from scratch using structured vibe coding principles, GitHub Copilot Coding Agent, and Azure AI Foundry. Instead of diving directly into prompts, I began with a specification which I wrote myself. It described:

What the tool should do
Which agents would be part of the system
What tools each agent could use (for example, Bing search grounding for the question answerer and checker)
How workflows should run, whether from a single question or an imported Excel file
UI preferences, illustrated with a screenshot

With that spec in hand, I handed it off to the coding agent. The first pass wasn’t exactly what I envisioned, but it gave me a working analog: tabs instead of stacked text boxes. It was different, but it worked, so I kept it.

This is the essence of structured vibe coding is to set the vibe through structure, provide agents with a clear blueprint, and let them improvise within those boundaries.

Turning GitHub Copilot into a Development Team

Of course, the first draft of my system wasn’t perfect. That’s where GitHub really shines: it allows me to treat my Copilot coding agents like developers on a team.

I began filing GitHub issues for every improvement I wanted, like:

Remove overly technical details from the reasoning log
Fix user experience issues in the main window
Make the answer checker less verbose
Clean up the output display

I assigned these issues to Copilot directly. Within moments, agents began working in parallel. I could track their progress through the GitHub Pull Requests (PRs) extension in Visual Studio Code, watching live as Copilot sessions unfolded.

When they finished, I had draft PRs waiting for my review. The changes were small but effective - removing unnecessary lines, tightening reasoning logs, and improving clarity. After running automated tests and scanning the diffs, I merged the PRs. Issues closed. The system improved.

This cycle - spec → issues → pull requests → review - felt familiar. It was classic software engineering, only now my “team” was comprised lf AI agents.

Watching the agents at Work

One of the more interesting moments came when I had three coding agents working on different issues simultaneously. Watching both of the pull requests evolve side by side was like managing a distributed engineering sprint.

The process was transparent. I could open one of the PRs, inspect its reasoning, review the code, and decide whether to merge. The experience underscored an important point: AI agents don’t replace the need for human oversight. Instead, they extend your capacity by taking on repetitive coding tasks so you can stay focused on direction and judgment.

Lessons Learned: Spec Driven Development

The key takeaway from this experiment is that managing AI agents is not fundamentally different from managing human developers. Both work best when you:

Start with a spec. Clear requirements give direction and reduce rework.
Use issues as a backlog. Structured tasks help break down work into manageable units.
Review pull requests. Oversight ensures quality and alignment with goals.

Without structure, it’s easy to fall into “prompt-and-pray” coding. You ask an agent to build something, get whatever comes back, and patch it manually. With structure, you create a development loop that feels natural, predictable, and scalable.

This philosophy echoes what the GitHub team recently described in their post on spec-driven development with AI. By starting from specifications, developers can make AI more than a novelty - it becomes a reliable part of the software lifecycle. And now, with GitHub’s introduction of Spec Kit, we’re making structured vibe coding part of the standard development process.

Structured vibe coding begins with a spec, and it continues with teams of agents fixing documented issues. This documentation is the crux of the structure the agents need to work properly.

In the end, I got a working app up and running in short order that has already saved me hours of manual effort. And I put it in this public GitHub repository so you can use it too: mcasalaina/QuestionnaireAgent_v2.

Looking Ahead

As AI agents continue to evolve, so will our approach to working with them. Structured vibe coding is just one way to blend human oversight with AI execution. It emphasizes what developers already know: specs and structure matter.

For me, the outcome was practical - I no longer need to copy and paste hundreds of questions into a tool by hand. But more importantly, the process itself provided a glimpse of what the future of software development may look like: humans and AI agents working together in a shared development environment, coordinated through familiar tools like GitHub.

The future is about equipping developers with AI teammates who can take on repetitive tasks, accelerate iteration, and free us to focus on design, direction, and impact.