Summary
Niklas Gustavsson, VP of Engineering at Spotify, joins Anthropic for an in-depth conversation about how the company is deploying AI coding agents at scale across one of software engineering’s most complex environments — a codebase exceeding 20 million lines of code, split between large monorepos and thousands of smaller polyrepos.
Gustavsson describes his personal workflow running five to ten parallel Claude Code sessions inside a tmux matrix with matching terminals, and details the architecture of Spotify’s internal agent system “Honk.” In its original version, Honk used an LLM judge to validate outputs and operated with a predefined allowlist of trusted tools. Honk V2, running on the Anthropic Agent SDK inside Kubernetes pods, has removed the judge — the models caught up — and allows users to bring their own tools. Critical capabilities include running CI builds on both Linux and macOS (enabling iOS simulator testing), and going directly from Figma designs to shipped UI implementations for Spotify’s TV app ports.
A major theme is the organizational investment required to make agentic development trustworthy at scale. As Spotify moved toward auto-merging PRs without human review, the company had to fundamentally strengthen test automation across thousands of component teams, each of which now owns their verification layer independently. Gustavsson also reflects on his personal “AGI moment” — the point, around Opus 4 or 5 in late 2025, when he realized he had stopped editing code manually entirely, marking the biggest shift in his 30-year engineering career.
📺 Source: Claude · Published June 29, 2026
🏷️ Format: Interview
