Coding With Devin: My New AI Programming Agent

Was this newsletter forwarded to you? Sign up to get it in your inbox.

We onboarded four new engineers at Every this week. 

The onboarding process was what you’d expect. I gave them access to our GitHub account so they could download our main code repos. I was peppered with the usual questions. 

“What’s the password to this account?” 

“Where do I find the API keys?” 

“Do you have any particular coding conventions you follow for this kind of task?”

There were some things, though, that were strikingly different about these new engineers. They welcomed being interrupted by me. They let me watch their computer screen in real time, so I could easily step in if they were making a mistake or got stuck. Also, they worked suspiciously hard. When I rolled over in the middle of the night to check my phone, I could see them still coding at 3 a.m.

Weirdly, they were all named Devin.

In case you have been living under a rock without access to X (if that’s you, congrats!), Devin is an AI software engineer. It’s an agent equipped with access to a full suite of developer tools—a browser, a command line interface, and an editor—as well as to your codebase. You can assign Devin small tasks like bug fixes or large tasks like prototyping an entirely new app from scratch. Devin will plan the work and keep you updated as it goes. You can watch along with everything it does and step in if you need to.

I’ve been using Devin for about a week, using it to do everything from fixing annoying little bugs that I haven’t had time for to building the first version of an app I’ve had in the back of my mind. Devin is almost impossibly futuristic and oddly familiar at the same time. I started with a few questions: Will it code for me so I can spend more time at the beach? Will it 10x Every’s engineering output at no extra cost? What does it mean for the future of software?

I’ll have a full review of Devin soon—with step-by-step details of how I used it and comparisons to other coding agents—but here are my first impressions.

Devin is oddly familiar

Working with Devin is familiar because it feels like adding a few junior engineers to your team. 

You toss them tasks, and they’ll get started with enthusiasm. To your mild surprise, code is getting written without your involvement. But it takes a long time to do the work, and they’re likely to get stuck every once in a while unless you’re keeping close tabs on them. They’ll often come back with a functional solution that, on closer inspection, isn’t optimal.

Just like working with junior engineers, working with Devin requires a skill set. You have to learn what it’s good at and what it fails at. You have to learn how to write tasks that it can understand and when to step in. You also have to learn how to actually step back and trust it to do the work, even if it won’t get the work done in the exact way you would’ve done it.

Hiring junior engineers doesn’t mean you get to spend all of your time at the beach. Instead, it means your time is spent delegating tasks and getting in the weeds as a pair programmer to help your reports get unstuck.

With junior engineers, this can feel like a drag, but with Devin, it’s a little different because Devin is almost impossibly futuristic.

Devin is impossibly futuristic

When I was in college, I became enamored with online poker. A good player can set up four, eight, or 12 tables at once on a split screen, and shift among them like a chess grandmaster playing a crowd at a park. 

To me, there was something incredibly sexy about this. Anyone can be calm, cool, and collected, checking pre-flop with an Ace-King in a game with their buddies. It’s another thing when you’re playing every shark on the internet at the same time on dual monitors with tens of thousands of dollars at stake, every hour. That’s really rock and roll.

Devin feels a little like the online poker experience. You’re coding, yes. But you’re coding four things at once in different tabs. And because each tab is its own development environment—with its own codebase, browser, and planner—it all feels manageable. It’s programming leverage; it’s productivity power.

Is it truly faster, today, than doing it yourself if you were willing to spend the time and focus? Probably not. But there are many tasks that you’d love to have done that you can’t devote your full attention to.

Are you getting to know the problems you’re solving nearly as intimately as you would if you were focused on one thing at a time? No. Is it suitable for problems that require that? Also no. 

But many programming tasks require the same amount of time and attention that a pro poker player needs to figure out what to bet. The rest of the time is just spent typing. 

It’s also entertaining, at least for now. It’s captivating to watch Devin plan a task, navigate the web for research, and write code. It’s the best agent experience I’ve ever seen—better than Github Copilot Workspace, which I reviewed a few weeks ago. I’m impressed by how good it is with today’s technology. Whatever’s going on under the hood is a significant step up from other agents I’ve used.

Devin and the allocation economy

I’ve been writing for a while about our shift from a knowledge economy—where you’re compensated based on what you know—to an allocation economy, where you’re compensated based on your ability to allocate the resources of intelligence.

In an allocation economy, we become “model managers” learning to scope and delegate tasks for models to do rather than doing them ourselves, in many cases. My thesis is that the same skills that human managers have to learn—how to delegate, when to get into the details, how to develop taste—become increasingly important and widespread in this new world.

Devin underscores this impression for me. It doesn’t spell the end of programming work or the disappearance of programming jobs. Instead, it shifts what programmers do. You spend more time prompting and reviewing work, and less time writing the code from scratch. But you can also do four things at once and skip some of the drudgery that you’d find boring anyway.

In that sense, Devin is less like an AI employee and more like a programming productivity tool. It’s a fancy new English-based programming language bundled with its own development environment that is good for a specific set of tasks, just like a programming language like Javascript or a framework like React is.

I wonder if it would be better positioned that way—as a language and a development environment instead of as an employee. Positioning Devin as an AI programmer gets programmers’ hackles up and raises expectations that Devin can’t possibly meet. For the foreseeable future, it’s not going to do everything without your involvement. Even first-time human programmers don’t often get coding tasks right when they start out.

But viewing Devin as a new programming tool might set the proper expectations.

There’s a learning curve with using Devin. It’s different from what you’re used to. It’s certainly only useful if you’re already technical yourself. But if you learn the ins and outs, you can significantly improve your productivity on certain tasks, like fixing minor bugs that have been bothering you, implementing self-contained features that no one else has time for, or building proof-of-concept prototypes.

I’m not surprised though, that Cognition—the makers of Devin—went with the AI employee angle over the productivity tool positioning. Even I couldn’t help but start this article with the AI employee angle—it’s too captivating. 

Even if it doesn’t do all of my work for me, Devin is very impressive for an early product. I’m excited to keep playing with it. More soon.