Microsoft is pitching a future where AI controls everything on your PC and agents go and do work for you in the background. But before the company gets there, it has to build the tools to make these systems work and convince its own developers that AI is actually capable of achieving these big promises.
Microsoft CEO Satya Nadella revealed earlier this year that up to 30 percent of the code of “some of our projects” is written by AI, and I’ve been eager to learn exactly how Microsoft’s developers are using the technology ever since. I’ve been speaking to sources and company execs to get a better idea of how AI is being used by Microsoft developers. Some employees have told me they’re skeptical that AI agents will be able to fully replace the work of humans, leaving developers to fix the mistakes of automated agents.
When I ask the company for more specifics, though, Microsoft touts its early success in deploying AI internally.
“We want to really look at where there’s developer toil, where we have inefficiencies,” says Amanda Silver, a CVP in Microsoft’s CoreAI team who leads product for the company’s Apps & Agents platform, in an interview with Notepad. “Part of what we’re looking at is both how we can apply [AI] and where we can apply it.”
There are over 100,000 code repositories inside Microsoft, from brand new projects to legacy codebases that are more than 20 years old and still up and running. “We have pretty much every programming language, architecture, and lifecycle stage that you can imagine, and this really reflects a lot of our customers,” says Silver. That’s a lot of code for AI to potentially touch, especially as Microsoft pushes beyond simple code completion towards more automation with AI agents.
In May, Microsoft embedded a coding agent directly into GitHub Copilot, letting developers assign it work to do. The agent then goes off and creates its own development environment, runs in the background, and creates draft pull requests. “What we see is that developers save on average 30 minutes on simple tasks, over a half day on medium tasks, and two weeks on complex tasks,” says Silver. Microsoft’s developers are using it for time-consuming and monotonous tasks like fixing bugs and improving documentation for apps and services.
Microsoft looks at developer hours saved, incidents mitigated, or estimated hours saved for tasks to get to these numbers. “Additionally, we look at the actions completed by the agentic capabilities, such as the number of pull-requests it contributes to,” says Silver.
Measuring the impact of AI on developer productivity is something that I hear Microsoft is obsessing about internally, even if some studies show AI can make experienced developers slower. Some employees, who wish to remain anonymous, feel that Microsoft executives aren’t happy with how often developers use AI right now. There’s a push internally to get developers to use AI first for everything, but I hear that adoption isn’t always organic.
“It does require a little bit of intentional engagement to allow the mindset shift to click in,” admits Silver. While Microsoft’s developers could ignore GitHub Copilot Chat because it was in a separate window, the agentic mode and coding agent are right in the context of how developers work. “It becomes habit forming and changes behavior,” says Silver.
Microsoft says 91 percent of its engineering teams use GitHub Copilot, but sources have shared data that suggests, in some parts of the company, overall AI tool adoption is much lower – closer to the 51 percent of developers who told Stack Overflow they’re now using AI tools professionally every day.
... continue reading