← Cases
SwedenMedia and entertainment2026 – ongoing

Spotify

Claude Code and the Honk agent for code migrations and daily engineering

Spotify's engineers had AI write 73% of pull requests, with Claude Code agents running in the background

WorkingVendor case studyVendor ↗

Problem

Spotify has about 2,900 engineers and a backend of more than 20 million lines of code. Large migrations and maintenance changes across thousands of repositories took months, and about ten were done each year.

Approach

Engineers ran Claude Code in terminal sessions, often with several background agents at once. Spotify also built Honk on the Claude Agent SDK to make migration and maintenance changes across repositories. A judge component checked the agent's output before a pull request was accepted.

Results

73%share of pull requests authored by AIVendor case study[1]
more than 75%improvement in pull request frequency, which Spotify attributed directly to AI toolingVendor case study[1]
80%success rate of Honk pull requests after an early judge component was added, up from roughly 20-30%Vendor case study[1]

Takeaway

Adding a check on the agent's output raised the success rate of automated pull requests sharply. Spotify said it was still working to link pull requests and deployments to user value.

Categories

Business function
IT and R&D
Innovation type
Process innovation
Evidence
Vendor case study

Sources

Source summaries

[1]Anthropic22 Aug 2026

Anthropic's Office Hours session with Spotify engineers covers how the company uses Claude Code and its Honk agent for daily work and large-scale code maintenance. It gives the share of AI-authored pull requests, the gain in pull request frequency and the success rate of Honk after a judge was added. The page carries no date; the Internet Archive first saved it on August 22, 2026.

Full details are in the original sources.

Added 4 Nov 2026 · Updated 4 Nov 2026
Open case page ↗