Stripe’s internal coding agents, called minions, write code that clears a weekly merge count above 1,000 pull requests on their own. Spotify has pushed over 1,500 AI-generated pull requests from a background agent named Honk into production. Amplitude tripled the pull requests it shipped over six months while its reported monthly bug total fell from 715 to 319. GitLab CEO Bill Staples cited all three in an August 24 essay on the company’s blog, arguing that the constrained resource in software engineering has moved: not producing code, but trusting it.
Staples frames the shift as another turn of a familiar cycle. Assembly language made hand-written machine code obsolete, and higher-level languages did the same to assembly. Each layer made one problem cheap and exposed the next one behind it. He argues large language models are repeating that pattern for code generation itself, while noting they are probabilistic systems rather than the deterministic transformations a compiler performs.
The supporting detail he cites comes from case studies those companies published themselves, not from GitLab’s own product telemetry. Stripe’s minions run in sandboxed compute kept warm for parallel jobs. Each must clear its own local checks first, then it is allowed to push and draw selectively from a suite exceeding three million tests. Amplitude cut the median span from opened pull request to merge from 5.2 hours down to 44 minutes, and shrank frontend CI from roughly 30 minutes to three or four, then used a risk model to decide which changes could merge without a person. Spotify built a verification layer that lets Honk check its own work without seeing how the verifiers themselves are built. None of the three, in Staples’s account, got dependable output just by writing a smarter prompt. Each paired the model with deterministic gates, isolated execution and an audit trail.
This is a company blog post, and GitLab sells the products built to be that layer: the Duo Agent Platform, a context graph called Orbit, and a governance tool for agent identity and approvals. Staples’s account of what Stripe, Spotify and Amplitude built is reporting on public material those companies chose to publish. His conclusion that enterprises need a vendor-neutral, model-agnostic platform to hold context and governance, rather than building it themselves, is a recommendation that happens to describe GitLab’s own roadmap. One number stands apart as a direct GitLab claim rather than a third-party result: the company says its rearchitected source control produced task execution up to 50 times faster in internal testing, a figure with no independent verification cited.
What an engineering leader can act on this quarter: Staples’s proposed metric, cost per accepted change rather than lines of code, needs no new platform to start tracking. Timing the CI pipeline is free too; a run well past five minutes is likely a bigger constraint than which coding model a team uses. Writing down merge criteria for one narrow class of low-risk change, the way Amplitude did for its SOC 2 approval workflow, is achievable without signing a vendor contract.
What requires organizational change no single team controls: a context and identity layer that survives a switch of model, cloud or agent vendor is a platform decision, not a team-level workflow fix, and it is the exact layer GitLab, Anthropic and rival platforms are now positioning to own. Treat Stripe, Spotify and Amplitude’s results as evidence the underlying problem is real. Treat GitLab’s specific architecture, described in its own blog post, as one vendor’s proposed answer to it, not the only one available.
GitLab published this analysis, written by CEO Bill Staples, on its company blog on August 24, 2026.
