Too Long; Didn't Watch — Summary
Stripe has transitioned to agentic engineering by deploying 'Minions,' custom unattended coding agents that ship 1,300+ pull requests weekly across a high-stakes, 100-million-line codebase.
Loading summary...
Loading summary...
4 min read (91% time saved)
Stripe has transitioned to agentic engineering by deploying 'Minions,' custom unattended coding agents that ship 1,300+ pull requests weekly across a high-stakes, 100-million-line codebase.
Summarize another video
Press ⌘K to quickly paste a new URL
Stripe engineers are currently merging 1,300 pull requests per week containing zero human-written code. Operating a system that handles $1.9 trillion in volume (1.6% of global GDP), Stripe cannot afford the risks associated with 'vibe coding.' Agentic engineering is defined as knowing exactly how a system will behave, whereas vibe coding is relying on AI without deep architectural control.
Stripe's system consists of several critical nodes:
Standard LLMs are not trained on Stripe's internal, homegrown Ruby libraries. Minions were created to solve the 'large codebase' problem where context is too vast for standard prompts. The goal is 'one-shot' task completion where an agent receives a Slack message and produces a production-ready PR. This allows for massive parallelization of developer attention.
Engineers interact with Minions via a custom UI that surfaces tool logs, thought processes, and modified files. This transparency allows for quick reviews. Once a task is done, the agent creates a branch, pushes to CI, and follows Stripe’s PR templates, requiring only a final human review.
Minions run on AWS EC2 instances that mirror human developer environments. These 'dev boxes' are pre-warmed to spin up in 10 seconds. This isolation prevents agents from damaging production systems and allows an engineer to run half a dozen agents in parallel without the overhead of Git worktrees.
Stripe forked the 'Goose' agent harness to create 'Blueprints.' Blueprints are workflows designed in code that direct a Minion's run. They use deterministic code for tasks like linting and git commits, while using agents for creative problem-solving. This hybrid approach is more cost-effective and less brittle than using agents for every single step.
To manage a repository with millions of lines, Stripe uses scoped rule files (similar to .cursorrules). These markdown files use glob patterns to conditionally load context only when an agent traverses specific subdirectories. This prevents 'token explosion' and ensures the agent only sees relevant documentation and standards.
Stripe manages nearly 500 Model Context Protocol (MCP) tools through a centralized 'Tool Shed.' This is a meta-tool that allows agents to discover and select the specific tools they need for a task. This 'meta-agentic' approach (agents building or selecting tools for other agents) is essential for scaling intelligence.
The system is rated an 8/10. Two areas for improvement are identified:
"Agentic engineering is knowing what will happen in your system so well you don't need to look. Vibe coding is not knowing and not looking."
"Specialization is how you win... your tooling and your code must also be specialized."
"You want to be building the system that builds the system... that's the lever that agentic engineering unlocks."
"There are many coding agents, but this one is mine."
IndyDevDan
4 min read
IndyDevDan
8 min read
IndyDevDan
7 min read
IndyDevDan
4 min read
IndyDevDan
7 min read
IndyDevDan
5 min read