Back office ops · Production

Spotify's background coding agent (Honk) generates 1,500+ merged pull requests across the codebase

The problem

Spotify's Fleet Management system handled simple, repeatable code transformations well but could not solve complex migrations. Defining source-to-source transformations programmatically required deep specialized expertise, and the complexity caused one migration script alone to grow to over 20,000 lines of code, limiting advanced use to a small number of teams.

Workflow diagram · grounded in source

Engineer defines task in natural language

trigger

“The goal was to allow engineers to define and run fleet-wide changes using natural language.”

Interactive agent gathers task context

ai_action

“They first talk to an interactive agent that helps to gather information about the task at hand. This interaction results in a prompt that is then handed off to the coding agent, which produces a pull request.”

Coding agent executes prompt

ai_action

“We replaced deterministic migration scripts with an agent that takes instructions from a prompt.”

LLM-as-judge diff evaluation

validation

“evaluate a diff using LLMs as a judge”

Pull request opened automatically

output

“automatically open pull requests against the target repositories”

Human team reviews and merges

human_review

“teams across Spotify have merged into our production codebase”

Agent config improvements feed all future PRs

feedback_loop

“improvements to the agent configuration and tools now apply to any PRs generated, no matter the source of the task”

Reported outcome

Spotify's AI coding agents have generated more than 1,500 pull requests merged into production, achieving a time saving of 60–90% compared to writing code by hand, while enabling complex migrations that were previously out of reach for most teams.

Reported metrics

AI-generated pull requests merged into productionmore than 1,500

Time saving vs writing code by hand60–90%

share of Spotify pull requests automatedaround half

Developers interacting with the agenthundreds

Reported stack

Fleet ManagementMCPGCPMLflowLLMsSlackGitHub Enterprise

Source

https://engineering.atspotify.com/2025/11/spotifys-background-coding-agent-part-1

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Spotify's AI coding agents have generated more than 1,500 pull requests merged into production, achieving a time saving of 60–90% compared to writing code by hand, while enabling complex migrations that were previousl…

What tools did this team use?

Fleet Management, MCP, GCP, MLflow, LLMs, Slack, GitHub Enterprise.

What results were reported?

AI-generated pull requests merged into production: more than 1,500; Time saving vs writing code by hand: 60–90%; share of Spotify pull requests automated: around half; Developers interacting with the agent: hundreds (source-reported, not independently verified).

How is this back office ops AI workflow structured?

Engineer defines task in natural language → Interactive agent gathers task context → Coding agent executes prompt → LLM-as-judge diff evaluation → Pull request opened automatically → Human team reviews and merges → Agent config improvements feed all future PRs.