I Took on a Hackathon at Microsoft Headquarters. How AI Agents Are Changing the Software Landscape

Last week, I participated in a hackathon held within Microsoft. It was an event connecting Microsoft headquarters with locations around the world. For one full week, I stepped away from my regular duties to focus entirely on hacking. Building, learning, and innovating. While I naturally use AI in my daily work to create prototypes and write documentation, this week was a time to focus solely on my own ideas and bring them to life.

Looking around, most people were participating in teams or groups. They divided tasks among those who build the UI, those who write the backend, and those who put together the presentations. I chose a solo project.

It is exciting to bring my own ideas to life while using AI to work efficiently every day.

Bringing Ideas to Life

I have IT appliances that I use at home all the time. I tried building a small app that could control those appliances from my PC. I based it on the Nature Remo Cloud API. Once you issue an access token, you can do everything from retrieving a list of devices to operating the air conditioner and TV with just HTTP requests. I knew that APIs and MCP servers for these existed, but I also had AI research which technologies to use and how to implement them. Saving time on research is truly helpful. The tool I used here was the GitHub Copilot CLI. One major, powerful advantage is that you can use various models. Learning about the characteristics of each model was also one of the goals of this hackathon.

The Power of GitHub Copilot CLI

I kept the Copilot CLI resident in my terminal and handled everything from implementation to debugging and checking API usage in one go. Without stopping my hands from writing code, I asked for the cause of errors on the spot, received suggestions for fixes, and moved on to the next step. Not stopping. That became the greatest weapon for my solo project.

I didn't even have time to hesitate over the choice of language or framework. As long as I could hit the Nature Remo API, the language didn't matter. This time, I chose the shortest route with a stack I was familiar with, and used the saved time for appliance control logic and UX.

If you think while interacting with an AI agent, that becomes your development speed. I felt like I could move at double speed even on my own. It becomes a battle of how far you can make tools follow your own thinking. What determines the limits of an individual is not the number of people, but the ability and speed of interaction with AI.

What Is Required for PM-Led Software Development

When AI agents start to handle the majority of implementation, the skill set required for PMs also changes. Beyond the judgment of what should be built, the design ability of how to interact with AI, how much to delegate, and where humans should make decisions is called into question. It is an era where those who can "use it to the fullest" are stronger than those who can "build it." This should become a central point of discussion in future product development.

Astra or Claude? If you want accuracy, this was the one

In this hackathon, while I proceeded with implementation via the Copilot CLI, I tried several models to run in the background. I compared OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5 by throwing the same tasks at both.

To conclude, if you want to build accurate software, Astra was the only choice. Although there were times when the response took time, its precision in picking up fine edge cases—such as SDK authentication and differences in API specifications for each device—was a cut above the rest. Honestly, Astra is the one I liked the most. Claude Opus 5 had a sense of speed and there were times when there was less rework, but for tasks like this where accuracy is directly linked to the success of the demo, there were many times I chose Astra's output even if it meant waiting a bit.

Claude for speed, Astra for accuracy. My honest feeling is that you need to use them differently depending on the purpose. I see it becoming the standard in the future to operate by switching models between the phase where you want to iterate with a focus on speed and the phase where you refine the final accuracy.

Looking Ahead

What I became convinced of through this experience is that the productivity of individual development will jump even further in the future. The areas where the very composition of team versus individual loses its meaning will expand. From the era of teams to an era where individuals hold weapons.

With a development flow centered on AI agents, an era is coming where my prototype creation directly leads to contributions to the company and my own impact.