How Spotify runs 5-10 AI agents in parallel on a 20-million-line monorepo

5-10 AI agents in parallel on a 20-million-line monorepo: Spotify’s tmux + git worktree setup

The setup: 5-10 Claude sessions in tmux, one per git worktree

How do people actually work with AI agents at enterprise scale?

The explanation comes straight from Niklas Gustavsson, VP of Engineering at Spotify.

In an interview with Boris Cherny, creator of Claude Code, he described his setup:

• 5-10 Claude sessions open at the same time in tmux
• one session for each git worktree
• agents left working in the background
• all inside a backend monorepo with over 20 million lines of code

He expected problems. Instead, by his own account, it works surprisingly well.

The numbers: 73% of PRs AI-assisted and 4,500 deploys a day

Spotify’s numbers are impressive:

• 73% of Pull Requests are AI-assisted
• PR frequency has increased by 75%
• around 4,500 production deploys per day
• 99% of engineers use AI tools every week

Setup and numbers come from the video interview published by Anthropic on June 29, 2026 and from the Spotify Engineering post of June 3, 2026.

I use the same pattern on a much smaller scale: AI agents on analysis, development, testing, and documentation, with technical control and the decisions always mine. It’s how I run custom software projects for companies.

The infrastructure behind the results

But it takes more than installing Claude Code to get these results.

Over the years, Spotify has built an infrastructure that makes the codebase understandable, standardized, and easy to modify.

1. Backstage

Backstage is the open source Internal Developer Portal created by Spotify.

It centralizes in a single interface:

• services, libraries, pipelines, and websites
• component ownership
• dependencies
• documentation
• system health status

In practice: instead of hunting for information across Slack, wikis, GitHub, and other tools, developers and AI agents can find everything in one place.

Why does it matter in the context of AI?

It provides standardization and structured context for the codebase. AI agents (like Claude) work much better when the code is consistent and well described.

2. Fleet Management and Fleetshift

Fleet Management is Spotify’s philosophical and technical approach to managing software at the “fleet” level (fleet = the set of all components), instead of component by component.

Fleetshift is the concrete tool that turns this vision into reality. What does it do exactly? It makes it possible to apply code changes across thousands of repositories at the same time.

Typical examples:

• Updating a dependency (e.g. a library) across hundreds of services
• Running complex migrations (e.g. from Java AutoValue to Records, framework upgrades, etc.)
• Applying security fixes or large-scale refactorings

Operations that used to take months can now automatically generate Pull Requests ready for review or, in some cases, for auto-merge.

3. LLM-as-judge

Every Pull Request is also analyzed by an AI model.

Thanks to this system, the success rate of agent-generated PRs went from 20-30% to around 80%.

4. Honk

Spotify has also developed Honk, a proprietary AI agent built using the Claude Agent SDK.

What really matters in 2026

The most important lesson?

In 2026 the real advantage comes above all from everything built around the AI agents: standardization, documentation, ownership, automation, and control systems.