NerdOut@Spotify

Welcome to Dev Mode, where we show you how Spotify builds, with live demos on our own repos and PRs. No slides. Just the workflows we actually use on real Spotify problems, plus an honest conversation about what’s working and what still trips us up.
In this episode, Dave Zolotusky is joined by Jonatan Dahl (JD), Staff Engineer, and Stuart Clark, Sr. Developer Advocate, to tackle a problem you’ve probably run into: you hand an AI agent a big task, and somewhere along the way it runs out of context, starts going in circles, and forgets what it was doing. So what if you stopped asking one agent to do everything?
JD shows us how he splits one big piece of work across a bunch of agent sessions in ⁠Xirp⁠, our mission control for AI coding agents, and keeps tabs on every one of them as it moves from in progress to merged. Then we get into the part that really makes or breaks it: context.
What does each agent actually need to know, what should you leave out, and what happens when you get it wrong? (Spoiler: things go off the rails pretty fast.)
Along the way, Dave, Stuart, and JD tackle the questions we all have:
When is one session enough?

When does splitting work into smaller PRs make life easier for the people reviewing it?

Which habits have actually stuck around at Spotify, versus the ones we quietly dropped?

If you’ve ever watched an agent lose the plot halfway through a task, this one’s for you.
Give Xirp a spin 👉 xirp.spotify.com

How do you build a culture of experimentation inside a 100-year-old institution?
This episode is part of our special miniseries on product experimentation where we go beyond A/B testing basics to explore how teams build experimentation culture, move faster with data, and make smarter product bets in practice.
Hosting this series are two of Spotify’s finest experimentation nerds: Luke Frake, group product manager for subscriptions, and Mårten Schultzberg, staff data scientist for Spotify’s experimentation platform, Confidence (confidence.spotify.com).
In this second episode, we explore what it’s like when your streaming content provider is also a national institution. As the UK's national public service broadcaster, the BBC produces everything from news and sports to drama and education across major digital services used by millions every day. With products spanning iPlayer, Sounds, and over a dozen mobile apps, experimentation at the BBC means optimizing for vastly different formats, platforms, and audience needs — all without the subscription metrics that most digital companies rely on.
Luke Frake and Mårten Schultzberg are joined by Divya Isaiah, Head of Experimentation at the BBC. She walks us through how the BBC runs a decentralized experimentation model supported by a center of excellence that acts as an in-house consultancy, filling gaps across teams at different levels of maturity. The conversation also explores how Divya approaches experimentation velocity (both the number of experiments and the number of teams experimenting), why the BBC's editorial teams are now getting into A/B testing, and where AI is showing up in experimentation tools versus where it's actually adding value.
Hear what happens when a venerable institution like the BBC embraces a culture of experimentation in order to inform smarter product decisions.
Learn more about how we experiment at Spotify:
Dive deep into experimentation nerdiness on the Confidence blog
👉 Here’s a good post to start with: “Beyond Winning: Spotify's Experiments with Learning Framework”

Get started with Spotify’s product experimentation platform: Confidence 

Become an experimentation expert with our free experimentation bootcamp

Follow the Confidence team on LinkedIn for product updates, event announcements, experimentation insights, and more

👋 Follow Luke on Linkedin

👋 Follow Mårten on Linkedin

Read what else we’re nerding out about on the Spotify Engineering Blog: engineering.atspotify.com

You should follow us on Twitter @SpotifyEng, LinkedIn, and YouTube!

What happens when a seven-person data team restarts an experimentation program from scratch — and gets the whole company running experiments within months?
Welcome to the first episode of our special miniseries on product experimentation. Now that you can build almost anything faster than ever, what matters most is learning what to build, what to keep, and what to put back on the shelf. Over the next few episodes, we go beyond A/B testing basics to explore how teams build experimentation culture, move faster with data, and make smarter product bets in practice.
Hosting this series are two of Spotify’s finest experimentation nerds: Luke Frake, group product manager for subscriptions, and Mårten Schultzberg, staff data scientist for Spotify’s experimentation platform, Confidence (confidence.spotify.com).
In this first episode, we head outside Spotify to explore how Huel, the UK-founded nutrition brand, approaches experimentation across a rapidly growing omnichannel business. Unlike pure digital products, Huel sells direct-to-consumer on its own site, on Amazon, in supermarkets, and on TikTok Shop — which means you can't always rely on a classical A/B test to measure what's working.
Luke and Mårten are joined by Bhav Patel, Data Director at Huel, to dig into how his lean team rebuilt Huel's experimentation culture from the ground up. They discuss why Huel lets experiment ideas come from anyone and how a founder-led, data-first culture drove strong adoption of experimentation across the entire org. You'll hear how Bhav's team tackles the unique challenges of omnichannel measurement: building synthetic control groups to evaluate TV ads, running geo-lift tests, and stitching together fragmented user data across platforms, plus how AI is fast-tracking experiment design and more.
If you care about building better products with data, check it out — and stay tuned for more episodes in this series dropping soon.
Learn more about how we experiment at Spotify:
Dive deep into experimentation nerdiness on the Confidence blog

👉 Here’s a good post to start with: “Beyond Winning: Spotify's Experiments with Learning Framework”

Get started with Spotify’s product experimentation platform: Confidence 

Become an experimentation expert with our free experimentation bootcamp

Follow the Confidence team on LinkedIn for product updates, event announcements, experimentation insights, and more

👋 Follow Luke on Linkedin

👋 Follow Mårten on Linkedin

Read what else we’re nerding out about on the Spotify Engineering Blog: engineering.atspotify.com
You should follow us on Twitter @SpotifyEng, LinkedIn, and YouTube!

What does it take to ship a video-first Spotify experience on the biggest screen in your house?
In this Release Notes episode of the NerdOut@Spotify podcast, you’ll hear about the redesigned Spotify app on Apple TV — why we rebuilt it, what it took to build a fully native tvOS app from the ground up, and how we evolved from a templated TVML-based UI to a modern, pixel-perfect experience.
The result is a faster, more visual experience tailored for a bigger screen, with expanded video support, AI DJ, on-screen lyrics, and improved navigation across the app.
Along the way, we dig into tvOS focus-based navigation, service runtime and dependency wrangling, and the important role AI coding agents played in helping us move faster — from generating UI from Figma designs, to mapping complex dependency trees, to automating visual diffs for pixel perfection.
The new Spotify app is available on the Apple TV App Store now.
Read what else we’re nerding out about on the Spotify Engineering Blog: engineering.atspotify.com

You should follow us on Twitter @SpotifyEng, LinkedIn, and YouTube!

At Spotify, we're doubling down on background coding agents. The goal is to scale code migrations and boost developer productivity beyond our existing automation system. Built on top of Fleet Management and supercharged by Claude Code, these agents make their own code changes and pull requests. What could possibly go wrong?
Engineers Jo Kelly-Fenton, Aleksander Mitic, and Max Charas join host Dave Zolotusky to unpack how we’ve merged thousands of AI-generated PRs while keeping codebases consistent and secure. They’ll dig into what it takes to let agents touch production code at scale: sandboxing and security, handling context windows, wiring into CI, and building verification loops that still trust humans more than agents.
You’ll hear real-world failure modes (like agents “fixing” tests by deleting them), how using an “LLM as a judge” flags sketchy diffs before they land, and how engineers and PMs are now triggering agents directly from Slack for migrations, documentation fixes, and feature flag rollouts. 
It’s a candid look at what increasing reliance on AI-based tooling means for the future of engineering — and for safely shipping code at speed.
Learn more about our journey with background coding agents:
1500+ PRs Later: Spotify’s Journey with Our Background Coding Agent (Part 1)

Background Coding Agents: Context Engineering (Part 2)

Background Coding Agents: Predictable Results Through Strong Feedback Loops (Part 3)

AI Engineer Paris 2025: Rewriting all of Spotify's code base, all the time

Read what else we’re nerding out about on the Spotify Engineering Blog: engineering.atspotify.com
You should follow us on Twitter @SpotifyEng, LinkedIn, and YouTube!