How Google’s dev tools manager makes AI coding work

As Google’s Project Manager for Developing Tools, Ryan Salva has a front row seat as AI equipment is changing coding. Previously Github and Microsoft, he is now responsible for the tools like Gemi CLI and likes Help the Gemini CodeDevelopers’ nooding in the new world of agent programming.

His team has been released New third party research Tuesday shows how developers actually use AI tools – and how much progress remains. I sat down with Salver to report to the AI ​​coding tools and talk about his personal experience.

This interview was edited for length and precision.

Every year, Google does a survey of trends of developers – but this year’s report really focuses on AI equipment and especially how agent developers are interested in programming. There was something in the study that made you surprised?

One of the most interesting searches was the middle date when developers started using AI tools. They found that it was April 2024, which came out of Claud 3 and the Gemi 2.5 came out fairly neat. This is really the dawn of logic or thought models and at the same time, we have found better in the equipment-coling.

For coding tasks, you need to really be able to earn external information to solve the problem, so it may be necessary to gram it, it may be necessary to compile the code. If the code is compiled, it may seek that unit test and to conduct that integration test. I think that equipment-coling is truly important pieces that have given models the ability to self-repair as they move forward.

How are you personally using AI Coding Tools?

For these days for most of my coding hobby projects and I spend most of my time using the command line-based tools. So it includes the Gemi Klie. Then there is some closed code, there are some codes. And you never use the terminal-based equipment yourself, so I really do different around the ID I use. I use Z. I use the VS code. I use cursor. I use Windsoroff, all of them, because I am eager to see how the world works and how the industry is developing.

From the professional direction, the product directors live in documents, so the first thing is to help me write specifications and requirements dox to help me write.

I am curious about how it works. If you are using Gemi CLI to create Gemi CLI, but I will imagine that it just does not run itself.

A development work will usually begin as a problem, perhaps it is a githab issue that someone has dropped with bugs. Often, if I’m really honest, this is a fairly short-specific problem. So I will use Gemi CLI to create more views in Markdown. It will usually produce about 100 lines of fairly technical, but also the result-powered specification. Then I will use Gemi CLI to write the code based on general preferences in that specification and group documents.

Throughout the engineering team, we have different types of rules and marrowdown dox that are consumed by models, just determine the way to work: Here we test how we conduct dependence here. So when it creates the code, it also works from those documents.

And as Jemi CD is going on and the problem solve is working, I will update my requirement dock and say, “I have fixed this step now I’m in the next step,” and so on. Each of these can create its own promise and pull the request to the repository, so I can always rewind or undo.

I would probably say that 70% to 80% of my work I am working in the terminal with the natural language, trying to use Gemi CLI to create requirements and then Jemi CLI will write most codes for me, which I will then review and read what IDE I am using. However in most cases I am using IDE as a place to read the code instead of writing the code.

Do you think the raw computer code has the future? Or will we simply move all things to the terminal window?

For three decades, there was IDE where we went to do everything in the development of software. You had IDE, you had a browser and your terminal window was.

I think it’s still basically case, but I suspect that over time we will spend a lot more time to work with requirements and the time spent on IDE will gradually shrink. And I think that the change can actually happen on the horizon of a long time.

There is a lot of resentment about what it means to develop software as progress. If 10 years from now, we are no longer looking at the code, what does it mean for developers? Will they still have any work for them?

I think that as a developer, your job is going to look like many others like architects. It is about to take big, complex problems and break them in small, solved functions. To reveal it in the machine code you need to think like a bigger image about what you are trying to produce than the intermediate language.