Descriptions:
Stripe is now shipping over 1,300 fully AI-written pull requests every week — out of roughly 8,000 total — using an internal agent harness called Minions. Cole Medin digs into Stripe’s published engineering blog posts to unpack the architecture behind this system, explaining what makes it reliable enough for a company processing over $1 trillion in annual payment volume on an uncommon Ruby codebase filled with proprietary libraries.
The core design principle is the separation of deterministic and non-deterministic workflow steps. Rather than letting an AI coding agent own the entire process from planning through validation, Minions enforces guaranteed checkpoints — linting, type checking, and unit tests — after the agent writes code, automatically looping back with failure context when issues arise. The system controls the agent, not the other way around. Stripe also built a centralized MCP server called Tool Shed containing around 500 tools for internal systems, with a context curation stage that selects a relevant subset before any agent session begins.
Medin draws deliberate comparisons to similar frameworks at Shopify (the open-sourced Roast engine), Airbnb (test migration tooling), and AWS, arguing that structured AI workflow engines are becoming standard practice at scale. The video concludes with practical guidance on how smaller engineering teams can implement the same deterministic-plus-agentic node pattern without Stripe’s headcount or infrastructure.
📺 Source: Cole Medin · Published March 12, 2026
🏷️ Format: Deep Dive
