Stripe Says AI Now Writes Thousands of Code Updates Weekly, But Humans Stay in Charge
Stripe’s AI agents now generate over 1,300 weekly code updates, yet human engineers still review every single change.
Payments giant Stripe has revealed just how deeply artificial intelligence is embedded in its software development process. The company says its internal AI system now generates more than 1,300 pull requests every week that are entirely written by AI, without any human-authored code. However, Stripe emphasizes that human engineers still review every single change before it is merged into production systems.
The details were shared through Stripe’s engineering blog and social media posts, where the company offered a closer look at its AI coding agents, internally called “minions.” These AI agents are designed to operate independently. Once assigned a task, a minion can write the required code, run tests, address common issues, and submit a pull request without ongoing human supervision.
Stripe describes minions as fully unattended AI coding agents capable of completing tasks end to end. There is no engineer guiding each step in real time. Instead, the AI works within Stripe’s existing developer environment, known as a devbox — a cloud-based setup already used by human engineers.
Each devbox contains the necessary codebase and development tools, and importantly, it is isolated from sensitive systems and real customer data. Because these environments are easy to spin up and replace, Stripe can safely deploy multiple AI agents simultaneously. Engineers can assign different tasks to different minions, much like distributing work among team members.
Interestingly, Stripe notes that devboxes were originally built to improve human productivity. Now, those same tools have become foundational in enabling AI-driven development at scale.
Despite the high level of automation, Stripe makes it clear that humans remain central to the decision-making process. Every AI-generated pull request is reviewed by an engineer before it is accepted. While AI can efficiently follow instructions and resolve known issues, Stripe acknowledges that it still struggles with broader architectural thinking and nuanced judgement.
Complex considerations such as long-term system design, security implications, and unforeseen edge cases continue to require human expertise. In other words, AI may accelerate development, but it does not replace strategic oversight.
To ensure reliability, Stripe has introduced structured “blueprints” that guide how minions operate. Certain steps — such as running linters or pushing code — follow fixed, predictable processes. Other stages allow the AI more flexibility in determining solutions. This hybrid model helps reduce errors while preserving adaptability.
Stripe has also implemented safeguards to prevent inefficiencies. AI agents are typically limited to one or two attempts to fix failing tests. If issues persist, the task is handed back to a human engineer. This approach helps control costs and avoids endless retry loops.
According to Stripe, its experience demonstrates that AI adoption works best when built on strong existing engineering foundations. Years of investment in tools designed for human developers have enabled the company to integrate AI into its workflow more smoothly and safely.
For now, Stripe’s message is clear: AI may be writing much of the code, but humans are still firmly in control of what ultimately goes live.
