Spotify

Spotify publishes its internal agent tooling on engineering.atspotify.com, and that habit is why the company keeps reappearing here. On 3 September 2026, Dimitri Mazmanov, Principal Product Manager at Spotify, used that blog to describe a setup that cut his «Claude Code» token consumption by roughly 90% across four scenarios tested against a Java monorepo. The machinery is Spotify's own: the AiKA Modes of «Portal by Spotify», declarative agents on an ephemeral runtime defined by instructions, a model, parameters and «MCP» tools, with no infrastructure or API keys to manage. «Honk» comes out of the same organization.

Justin Reock, Deputy CTO at DX, names Spotify when listing companies that aimed at a real bottleneck instead of coding time: the company speeds incident resolution by pushing context automatically to its SREs. Bill Staples, CEO of GitLab, treats Spotify's experiments alongside Stripe's and Amplitude's as evidence of where constraints reappear once generation gets cheap, namely environment, CI, review and governance.

Three different vantage points, three different parts of the chain. What holds them together is a company that ships the tooling and then writes up what failed: Mazmanov's own post devotes a section to what the delegation cannot do, including a thread-safety bug the cheaper model missed. His conclusion is that the saving matters less than turning model routing into a configuration problem rather than systems engineering.