Thomas Dohmke

I still need to understand what that actually does, right?

Like, how do I decide whether I can merge this or not?

Of course, I can YOLO it, right?

It starts with the planning process.

You know, the process of describing what you actually want to implement and the key new skill that engineers, product managers, designers, whoever's using, you know, this orchestra of agents that helps us through our day, you will learn and internalize

What is the right level of description that the agent needs to do actually what I want it to be?

Now, of course, you can, you know, when you try out an agent like this, you can just give it a very abstract task and see what it does.

And very quickly, we'll find out.

And you mentioned earlier image generators.

I think image generators are actually the perfect abstraction of that concept, right?

Like the first time you used an image generator and you told it you want to render an image of puppies.

or a computer on a meadow or something like that, right?

You most likely didn't get exactly what you wanted.

You got something that was close to the description and then you knew that, okay, I write another prompt and you try to figure out how do I change my prompt to now get closer to the image I want to get to.

And the same is true for coding agents, right?

You will try it out.

You'll see, okay, it wrote something, but that's actually not what I wanted or it doesn't align with our standards here at GitHub.

And so then you go back to your description and adjust your prompt.

And as you learn that skill, you're going to get better on describing exactly the task that the coding agent can take, you know, run 45 minutes, submit a pull request, you review the code, and you're like, yep, this actually looks good and it does exactly what I want.

And you will be wrong at times.