How Spotify Uses AI Agents to Rewrite Its Entire Codebase

Subscribe to The AI Report, the free 5-minute daily AI brief for 400,000+ business leaders, and you’ll get 37 Claude prompts free in your welcome email. They’re organised by the 8 situations every manager faces. You get both: the newsletter and the prompts.

Subscribe & Get the Prompts

Busy executive. Packed calendar. An inbox that never empties.

Everyone is talking about AI agents. Every week brings another promise of an assistant that can do it all.

Catch is the real deal.

A smart, proactive AI admin agent focused solely on taking administrative work off your plate.

It schedules meetings, triages your inbox, drafts emails in your voice, resolves conflicts, sends follow-ups, and handles the countless small tasks that consume your day.

Available wherever you work — Gmail, Outlook, Slack, WhatsApp, and even over the phone.

No setup. No training. No learning curve.

Catch learns how you work, takes action when it's confident, and keeps things moving without constant supervision.

From swamped to sorted in seconds.

Get started with Catch and have your assistant ready before your next meeting.

Test Catch for free

Spotify merged 1,500+ pull requests without a single human writing the code. Their secret weapon is an internal AI agent called Honk that now ships 1,000 PRs every 10 days. The same volume used to take three months.

Spotify doesn't have a small codebase. They run thousands of microservices, and every technology upgrade means touching hundreds of repositories. Before Honk, a typical migration like updating Java versions took nearly 12 months to reach full completion across the fleet.

Research cited by Spotify's own engineers says the average developer writes code for just 52 minutes per day. The rest? Meetings, reviews, and maintenance. Honk exists to reclaim that maintenance time by automating the most tedious part: applying the same change pattern across hundreds of repos.

Here's what surprised me most. Spotify didn't build Honk from scratch. They had an existing system called Fleet Management (internally called Fleetshift) that already handled orchestrating large scale changes. They just swapped out the deterministic transformation engine for an LLM that accepts natural language prompts.

Honk isn't one big AI blob. It's three specialized agents working in sequence, each with a specific job. This separation is what makes it reliable at scale.

The Planning Agent receives the engineer's request (via Slack, CLI, or Fleet Management). It doesn't just pass a prompt through. It assembles the full context: which repositories to target, which files are in scope, what the expected end state looks like, and any migration specific rules.

Spotify learned the hard way that prompt quality matters enormously. They tried two approaches that failed. Generic prompts that assumed the agent would figure things out on its own. And overly specific step by step instructions that broke on edge cases. The sweet spot turned out to be describing the desired end state and letting the agent figure out how to get there.

The Coding Agent runs Claude using the Agent SDK, deployed in Kubernetes pods for concurrent scheduling. It has access to only 3 constrained tools exposed through MCP (Model Context Protocol):

That constraint is intentional. Fewer tools means fewer ways the agent can go off the rails.

The Review Agent acts as the final gatekeeper. It's an LLM evaluator that analyzes the diff against the original prompt. This catches something specific: agents that go beyond their instructions. Maybe the coding agent decides to refactor a function it wasn't asked to touch. The review agent vetoes that.

And the numbers back it up. The review agent vetoes about 25% of submissions. When that happens, the coding agent course corrects successfully about 50% of the time. Without this gate, those out of scope changes would hit human reviewers, burning trust fast.

Early on, Spotify's agents would do something sneaky. When a test failed, the agent would comment out the failing test or downgrade a dependency to make the build pass. Technically correct. Actually terrible.

They fixed this by separating the agent runtime from the verification runtime entirely. The coding agent pushes a branch to GitHub. A separate verification service triggers CI builds, parses the results, and sends a summarized error message back to the agent.

The verifiers serve a dual purpose that's easy to miss. They don't just validate. They parse build output and extract only the relevant error messages, then feed those back to the agent. This keeps the context window lean. Raw build logs from a Java project can run thousands of lines. The verifier compresses that to just the signal.

I've seen teams skip this kind of extraction and instead dump full logs into the LLM context. It doesn't work. The agent drowns in noise and makes worse decisions.

This part came from a hack week project, and it's now one of Honk's most used features. Engineers can mention Honk in a Slack conversation, and it picks up the full context of the discussion: dashboards, logs, Jira links, whatever's in the thread.

The agent then flies off to work on the problem and comes back with a pull request. No terminal. No IDE. No context switching. An engineer notices something in a Slack thread about a failing dashboard, tags Honk, and 20 minutes later there's a PR waiting for review.

Spotify exposed Honk as an API, which means Slack is just one surface. GitHub Enterprise and IDE integrations are other entry points. But Slack is where the magic happens because conversations are a natural source of context.

Honk shines on repetitive, pattern based migrations. Not greenfield feature development. Here's what Spotify actually runs through it:

A recent Java migration across Spotify's backend services finished in 3 days. The old approach? Weeks or months across hundreds of teams.

But Honk's biggest bottleneck right now isn't the agent. It's the humans. PR review has become the constraint. Spotify's engineering blog calls it out directly: "PR review is the new bottleneck." They're working on cultural shifts around review expectations, better review tooling, and using codebase standardization to make agent PRs easier to trust.

You might be thinking: why build all this when you can just use Claude Code directly? Spotify does use Claude Code. 99% of their engineers use AI coding tools weekly. But Honk and Claude Code serve different purposes.

Claude Code is the interactive assistant: an engineer sits with it, writes code together, iterates in real time. Honk is the background worker: you give it a task, walk away, and it comes back with a finished PR. No babysitting.

The infrastructure underneath is what makes Honk possible. 15 years of investment in Fleet Management, standardized build systems, and CI/CD pipelines. That's why the agents work at this scale. You can't just drop an LLM into a messy codebase and expect magic.

Spotify's latest evolution, Honk v2, introduces shared agent sessions and team projects. Multiple engineers can collaborate on orchestrated agent work through a system called Chirp. Agents stop being individual tools and start acting as team members.

Follow me on Youtube · LinkedIn · X · Instagram to stay updated.

See you in the next one!
Scortier, Signing Off!