Stripe engineers have built an internal system called Minions that automatically processes over 1,300 pull requests weekly without any human-written code. These unattended AI agents work completely independently, receiving tasks through Slack, executing them in isolated cloud environments, and delivering finished pull requests that have already passed automated tests. The system represents a fundamental shift from traditional coding assistants.

Unlike tools like Cursor or Claude Code that require constant human oversight, Minions operate autonomously using a blueprint system that alternates between deterministic steps and agentic loops. This approach leverages Stripe's existing infrastructure - devboxes that engineers already use for isolated development. The agents spin up in under ten seconds, work in QA environments isolated from production, and handle everything from running linters to pushing branches according to company standards.

The key insight isn't the AI model itself but the infrastructure Stripe built years before LLMs existed. Their massive codebase of hundreds of millions of lines, mostly Ruby with Sorbet typing, moves over $1 trillion in payment volume annually. This makes the unattended agent approach particularly valuable for handling overnight issues or routine fixes that would otherwise interrupt engineers' focus on complex problems.