If you’re anything like me — you can think back to countless memories of coding sessions running deep into the night.
Darkened room, everyone else is asleep, and it’s just you and that black screen in front of you.
Yesterday, was just like so many of those sessions.
I had just gotten back from a work trip and my mind wouldn’t stop running. I opened my computer — ready to make use of the late-night hours.
There was, however, one enormous difference. A difference that may persist for every coding night I’ll ever have in the future.
For the first time ever — code was being written from my computer and commited into my codebase — but not by me.
I was a spectator.
Devin, Cognition Lab’s “Junior Developer,” was doing all the work. It’s one of the many new AI code-generation engineers: Cursor, Poolside, Codeium, Magic Dev, Github Co-Pilot. But with a $175m raise at a $2 billion valuation and previously only Scott Wu’s childhood mathlete videos for the public to go on, Devin had been the most highly anticipated.
Devin launched yesterday with a $500-a-month price tag.
The steep price was a bit shocking. Cursor is $20 and has become my most used and loved companion.
But after working with Devin, it’s very clear what the Cognition team is going after. The focus isn’t on making you more efficient; rather, it is how it can replace 75% of you. And by ‘you,’ I mean the population of engineers.
I don’t mean to sound sci-fi, but it was hard last night not to feel a shift in how software is built and maintained. To be clear, I believe the number of developers will keep growing, but the role of the traditional software engineer is changing.
I’ll spend the rest of the article diving a bit deeper into what I worked on and some findings —
Building with Devin
I set out to build some form of an automated personal assistant: an end-to-end interface that surfaces important emails I haven’t responded to.
A couple weeks ago, I wrote about switching our CRM from Affinity to Attio. And while Attio is the superior product in almost all capacities, the one thing it doesn’t have, and I do miss about Affinity, is their “Unanswered Emails” tab.
So — the goal was to replicate that feature using the Google Gmail API.
I began the project with authentication so the program would be able to read my emails. Devin struggled with this. And therefore, I struggled with Devin.
“~an hour in - bullish on Cursor, bearish on Devin.” — I messaged my colleagues. I gave up and wrote the authentication code myself using Cursor.
But once the authentication was set up, Devin started to cook. And it was, actually, pretty insane to watch.
I asked it to create a function to identify emails from the past week that I needed to respond to.
Followed by “Can you also make sure that the emails are in my "Primary" inbox?”
This was one of the most magical moments. I don’t know what I expected… but what I didn’t expect was it to spin up a browser inside its workspace, search, and analyze Google documentation on the internet.
All on its own: it opened Google, navigated through its developer docs, and wrote the correct implementation. Wild.
Analyzed the Gmail API documentation for the `users.messages.list` and `users.threads.get` methods to determine how to filter messages in the Primary inbox and track responses within threads
Next, I asked it to create a front end for me.
Below is a screenshot of the interface it spun up. What was crazy about this — is that it deployed the website. My usual workflow is to test locally from the command line and then manually deploy it.
But Devin wrote the front end, tested it locally, and deployed it into its own public link.
At this point, it was 3 a.m., and I was finally tired enough to get to sleep.
But in the morning when I hopped back online, so did Devin — and now, as I write this, Devin is working on connecting the code to a Supabase backend database. Adding a new functionality so I can “Ignore” an unanswered email for messages I don’t need to respond to, and that action will be saved for future sessions.
Final Thoughts
I’ve gone through a range of emotions in my first 12 hours with Devin.
Ultimately, I’m still getting used to it, but here are some high-level initial thoughts:
The ambiguity around what code Devin is writing, and the forced detachment from your codebase is still a bit uncomfortable. Cursor feels intuitive — you’re looking at a Text Editor with your code on the left-hand side and Cursor’s suggestions and chat on your right. It only updates scripts when you click “Accept.” Devin, on the other hand, is very under the hood. And, at times, feels less agile. It’s adding files you don’t know the names of… downloading packages that you’ve never used.
Devin acts as a stand-alone human rather than a coding companion. Cursor feels like it’s upgrading you, making you into a 10x engineer. Devin, almost as advertised, is more like you’re hiring an intern to do the work while you write emails or watch a Netflix show.
Devin seems to work at human speed. And for some tasks, it might even be slower. I imagine this will change over time, but for example, it’s been working on setting up a Supabase database and connecting the scripts for the past hour. I can do this in less than 10 minutes. This goes back to the “Junior Developer” pitch. Below is a screen of it taking 30 minutes to complete a task.
Devin automatically deployed my website, effectively hosting it — which is likely where they’ll generate a decent amount of revenue. It’s a playbook we’ve already seen with Replit: I launched a few apps with Replit agents, hosted them there, and our Replit bill jumped from $20 a month to $20 a week. I eventually redeployed on Google Cloud Engine to save costs, but for less technical users who don’t understand the pricing structure, they are more likely to stick with what’s convenient and eat the bill. I expect the same thing to happen with Devin-hosted apps.
I’m still figuring out where Devin truly shines — code fixes, bug bashing, refactoring, testing, front-end development, documentation, optimization? It is clear it’s not at the level where it can apply to YC as a technical co-founder. That said, it’s impressive to see the reports from Ramp using it and Devin actively contributing to Supabase’s codebase. It may be better for large organizations and code bases rather than spinning up side projects.
“Who will win” the AI engineer wars has been one of the main topics of 2024.
However, my takeaway from Devin is that it is extremely powerful and a massive achievement — but also that it is a very different product and offering than Cursor.
I’m more convinced that code generation may be so transformative and it is such a big (and valuable) problem space to tackle — that there may be multiple winners. I know that is a very “non-venture” take; most of the time, it is winner-take-all.
But after getting a feel for Devin, I can see a world where developers will continue to use Cursor and other tools as AI-companions, and Devin might actually slot in as a co-worker.
Either way, it sounds cheesy, but we’re living in such a cool moment. It is, without question — the golden age of engineering.
I’m an investor at Chapter One, an early-stage venture fund that invests $500K - $1.5M checks into pre-seed and seed-stage startups.
If you have any questions on the data, or if you’re a founder building a company, please feel free to reach out on Twitter (@seidtweets) or Linkedin (https://www.linkedin.com/in/jamesin-seidel-5325b147/).
Wild!