On February 10, Spotify co-CEO Gustav Söderström told investors on the Q4 2025 earnings call that the company’s most senior engineers “haven’t written a single line of code since December.” The system behind this claim is Honk, Spotify’s internal AI coding agent that merges 650+ pull requests into production every month. The claim landed alongside a 23% year-over-year R&D spending cut and record quarterly operating income of €701 million.
AI productivity gains, the narrative goes, are transforming software development at Spotify. But the real story isn’t Claude Code — it’s the decade of infrastructure Spotify built to make Claude Code work. And the timing of the announcement, four days after Spotify locked down its developer API citing AI risks, reveals what this is actually about: cost reduction dressed as innovation.
The Claim: “Haven’t Written a Single Line of Code Since December”
The exact phrasing matters. As TechCrunch reported, Söderström said: “Our most senior engineers say they haven’t written a single line of code since December.” Note the verb: say. This is reported speech on an investor call, not an audited technical claim.
He pointed to late 2025 as the turning point: “A lot of things happened in December, including Opus 4.5 coming out to Claude Code. And we crossed the threshold where things just started working.” Claude Opus 4.5 actually launched on November 24, 2025, though Spotify may have begun deploying it internally in December.
Meanwhile, CFO Christian Luiga reported Q4 R&D expenses of €290 million, down 23% year-over-year. The two data points — AI productivity miracle and steep R&D decline — were presented back-to-back to the same audience of analysts and institutional investors. That juxtaposition was not accidental.
How Spotify Honk AI Coding Actually Works
Strip away the earnings-call polish and Honk has a genuinely interesting architecture. Engineers submit natural language prompts through Slack describing the code change they want. Claude navigates the codebase, makes modifications, runs formatters, linting, builds, and tests inside sandboxed containers, then opens pull requests automatically. Anthropic’s published customer story reports 650+ agent-generated PRs merged monthly, with teams claiming 60–90% time savings on complex code migrations.
Tool access is deliberately restricted. The system uses Model Context Protocol (MCP) to expose three controlled interfaces: a Verify tool for formatters, linters, and tests; a Git tool with limited operations; and a Bash tool locked to a strict allowlist. An LLM Judge acts as a quality gate, vetoing roughly 25% of agent sessions — and agents successfully self-correct about half of those vetoed cases.
Spotify selected Claude Sonnet 4.5 as the default model for fleet-scale work — not the flashier Opus, but the model optimized for throughput and consistency. As their three-part engineering blog series documents, the team tested Goose, Aider, and a homegrown agentic loop before settling on the Claude Agent SDK in July 2025. Chief Architect Niklas Gustavsson put it plainly: “Claude has consistently delivered the strongest performance for large-scale code transformation work.”
The Infrastructure Prerequisite Most Companies Don’t Have
Here’s where the story gets uncomfortable for anyone hoping to replicate Spotify’s results. Honk sits on three infrastructure layers that took years to build. First: Backstage, Spotify’s open-source internal developer portal launched in 2020, which catalogs every component and standardizes ownership across thousands of repositories. In a quiet irony, Spotify’s most altruistic engineering decision — open-sourcing their IDP — enabled its most self-serving one.
Second: Fleet Management, a framework developed in 2022 for applying code changes across hundreds or thousands of repos simultaneously. Without it, you can’t safely automate what you don’t understand. Third: thorough test suites that let automated verification actually mean something. Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 due to escalating costs, unclear value, and inadequate risk controls. The infrastructure gap is why.
The Numbers: 650+ PRs per Month, But What Do They Represent?
Over 1,500 merged agent-generated PRs as of November 2025, with more than 650 per month and climbing. Since mid-2024, approximately half of all Spotify PRs come from automated systems — though that figure includes traditional automation, not just AI.
But look at the use cases: Java AutoValue to Records conversion, framework upgrades with breaking changes, configuration schema transformations, UI component migrations. This is maintenance work — critical, time-consuming, and exactly the kind of repetitive code transformation where AI agents excel. An individual developer shipped 259 PRs in 30 days using Claude Code, and the work that scales best is the same everywhere: migrations, refactors, dependency bumps.
Spotify shipped 50+ features in 2025, but no public data distinguishes how many were Honk-conceived versus Honk-implemented. An AI that executes a migration 60–90% faster is a productivity tool. An AI that designs the next Prompted Playlists feature from scratch is something else entirely. Söderström’s earnings call blurred that line — and investors seemed happy not to ask.
The Contradiction: AI-Powered Internal Development, AI-Restricted External Access
On February 6 — four days before the earnings call — TechCrunch reported that Spotify announced sweeping restrictions to its developer API. Starting February 11, developers need a Spotify Premium subscription just to access Development Mode. Test users were slashed from 25 to 5. Multiple endpoints were deprecated outright.
Spotify’s official justification: “Over time, advances in automation and AI have fundamentally altered the usage patterns and risk profile of developer access.” As Musically reported, the restrictions are creating a trail of dead applications — genre visualization platforms, artist discovery engines, mood-based playlist tools. All eliminated because Spotify decided AI poses too great a risk to its data.
The platform power dynamic is stark. AI is a tool when Spotify wields it internally to slash R&D costs. AI is a threat when external developers use Spotify’s data in ways Spotify doesn’t control. On the same earnings call where Söderström celebrated AI transforming engineering, he framed it as the reason to lock the door on everyone else.
Full-year R&D fell 6% to €1.39 billion while operating income grew more than 50% and net income tripled to €1.17 billion. Those AI-driven efficiency gains flow to Spotify’s bottom line — not to the royalty pool. Artists and labels don’t share in the savings.
Industry Context: The Infrastructure Gap Is Real
Spotify isn’t operating in a vacuum. AI writes roughly 30% of Microsoft’s code and over 25% of Google’s, according to the heads of those companies. Mark Zuckerberg has said he wants most of Meta’s code AI-generated in the near future. MIT Technology Review named generative coding one of its 2026 breakthrough technologies on January 12.
But the counter-evidence is equally real. The METR study from July 2025 found experienced developers were actually 19% slower with AI tools on familiar codebases — a result that stunned the participants, who predicted they’d be 24% faster. The difference? Infrastructure.
Spotify’s engineers aren’t wrestling with AI in random codebases. They’re deploying Claude into a system specifically built to tell the model what it’s looking at, what tests to run, and what the expected output looks like. That context engineering — the trust problem most teams haven’t solved — separates a 60% time savings from a 19% time loss.
The Question Nobody’s Asking
If senior engineers are supervising AI instead of writing code, what are junior engineers doing? How do you develop the engineering judgment required to review AI-generated pull requests when the code writes itself? Spotify hasn’t addressed this, and neither has anyone else deploying AI agents at scale.
The mentorship pipeline that turns junior developers into senior ones assumes those juniors spend years deep in code. If that pipeline breaks, who reviews the AI’s work in five years?
Spotify’s real innovation isn’t Honk. It’s the unsexy infrastructure — component catalogs, cross-repo frameworks, thorough test suites — that they spent half a decade building before AI agents existed. Everyone wants the Claude Code productivity numbers. Almost nobody wants to spend five years building Backstage first. That’s the gap Gartner’s 40% cancellation forecast is measuring, whether it knows it or not.
Spotify’s Q1 2026 earnings arrive in late April. Whether “haven’t written code” becomes a sustained claim or gets quietly retired as a one-quarter talking point will reveal whether this is genuine transformation — or the best-packaged cost-cutting narrative in tech.
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.
