Spotify Says Its Best Developers Haven't Written a Line of Code Since December, Thanks To AI
10 103Spotify's best developers have stopped writing code manually since December and now rely on an internal AI system called Honk that enables remote, real-time code deployment through Claude Code, the company's co-CEO Gustav Soderstrom said during a fourth-quarter earnings call this week.
Engineers can fix bugs or add features to the iOS app from Slack on their phones during their morning commute and receive a new version of the app pushed to Slack before arriving at the office. The system has helped Spotify ship more than 50 new features throughout 2025, including AI-powered Prompted Playlists, Page Match for audiobooks, and About This Song. Soderstrom credited the system with speeding up coding and deployment tremendously and called it "just the beginning" for AI development at Spotify. The company is building a unique music dataset that differs from factual resources like Wikipedia because music-related questions often lack single correct answers -- workout music preferences vary from American hip-hop to Scandinavian heavy metal.
10 comments
Guess who'll be kicked to the curb real soon (Score: 5, Insightful)
by haruchai ( 17472 ) on Friday February 13, 2026 @02:33PM (#65987190)
and they won't get a Honkin' big severance either
Re: Guess who'll be kicked to the curb real soon (Score: 5, Insightful)
by ArmoredDragon ( 3450605 ) on Friday February 13, 2026 @03:59PM (#65987386)
Until they release a breaking bug on their morning commute, and some people on the bus suddenly lose their Spotify. Then said developer has no idea what went wrong, likely for hours or even days, while the AI keeps hallucinating fixes that don't work, as both it and the developers have no idea what they're actually doing.
Please don't (Score: 5, Insightful)
by OrangeTide ( 124937 ) on Friday February 13, 2026 @02:37PM (#65987208)
Please don't work during your morning commute. Especially if you're the one driving.
But almost as importantly, if your employer makes you come into the office then you should ONLY work while at the office. And they can go F themselves if they want you to work on your own time as well.
Re:Please don't (Score: 5, Insightful)
by 93 Escort Wagon ( 326346 ) on Friday February 13, 2026 @02:51PM (#65987242)
Maybe I'm just naive* but - it's hard for me to imagine a competent developer willingly allowing new code to be "pushed to Slack" before they have a chance to run through the changes with their own eyes.
This doesn't pass the smell test.
* I realize this may be true regardless
Is it true? (Score: 5, Interesting)
by Revek ( 133289 ) on Friday February 13, 2026 @02:38PM (#65987210)
Is it true that AI code can't be copyrighted? If so, could spotify's code now be freeware? Semi trolling minds want to know.
Re: Is it true? (Score: 5, Funny)
by broward ( 416376 ) on <browardhorne.gmail@com> on Friday February 13, 2026 @02:52PM (#65987246)
this story reminds me a company meeting where our new manager was introduced as having delivered his last two projects with zero bugs.
we all burst into laughter at the same time
followed by an awkward silience.
AI Hype needs money (Score: 5, Informative)
by fuzzyf ( 1129635 ) on Friday February 13, 2026 @02:47PM (#65987230)
This smells like bs to me. No way experienced developers are letting AI generate bug fixes or entirely new features using Slack to talk to AI on the way to work.
The only question here is: What are they selling?
Increased stock value?
AI Coding tool that management has a stock option for?
The simple fact is that AI can generate code, but has absolutely no understanding of anything. It's a very useful tool, but not as what this bs article is trying to sell it as.
Re:AI Hype needs money (Score: 5, Interesting)
by Brain-Fu ( 1274756 ) on Friday February 13, 2026 @03:22PM (#65987290)
The experiences reported in these articles are so utterly unlike the ones I have using AI to generate code. It HAS gotten better in the last year, but it is still no where near this capable, for me.
If I give it too many requirements at once, it completely fails and often damages the code files significantly, and I have to refresh from backup.
If I give it smaller prompts in a series, doing some testing myself between prompts, there is usually something I need to fix manually. And if I don't, and just let it successfully build on what it built before, the code becomes increasingly more impenetrable. The variable names and function names are "true" but not descriptive (too vague, usually) and when those mount up the code becomes unreadable. It generates code comments but they are utterly worthless noise that point out the outright obvious without telling you anything actually useful. When new requirements negate or alter prior ones, the AI does not refactor them into a clean solution but just duplicates code and leaves the old no-longer-needed code behind and makes variable names even more weird to make up for it. The performance of the code decays quickly. And on top of all this, it STILL can't succeed at all if you need to do anything that is a little too unique to your business needs. Like a fancy complex loose sort with special rules or whatever. It tries and fails, but tells you it succeeds, and you get code that doesn't work.
Sometimes it can solve surprisingly hard problems, and then get utterly stuck on something trivial. You tell it what is wrong and it shuffles a lot of code around and says "there, fixed" and it is still doing exactly what it did wrong before.
I have good success getting new projects started using AI code generation. When it is just generating mostly scaffolding and foundational feature support code that tends to be pretty generic, it saves me time. But once the aspects of the code that are truly unique to the needs start coming into focus, AI fails.
I still do most of my coding by hand because of this. I use AI when I can but once this stuttering starts happing I drop it like a hot potato because it causes nothing but problems from then on.
I simply don't see how the same solution could reliably make consistent and significant changes to a codebase and produce reliable, performant, or even functional code on an ongoing basis. That hasn't ever worked for me and still doesn't, even with the latest gen AI models.
why commute? (Score: 5, Insightful)
by awwshit ( 6214476 ) on Friday February 13, 2026 @03:18PM (#65987284)
If you can do all of your work while commuting then why commute at all? Obviously does not matter where you sit.
Every threat actor now aiming for slack (Score: 5, Interesting)
by MNNorske ( 2651341 ) on Friday February 13, 2026 @04:02PM (#65987396)
If your AI can act on instructions given in slack, update code in source control, and then compile/deploy that code you just opened a whole can of worms. If I were a threat actor I would 100% be aiming to try and compromise their slack. Just tell the AI to introduce these few lines of code into the build... Or add this feature... It sounds like a security nightmare to me.