Stripe’s 1,300 PRs/week
Stripe’s internal AI “minions” are generating and merging more than 1,300 pull requests per week from Slack emoji reactions — tightly integrated with CI/CD, logging, and human review. That workflow is being cited as a template for agentic automation in engineering orgs, raising new questions about permissioning, audit trails, and rollback design. (lennysnewsletter.com)
Stripe built Minions as “one-shot” coding agents that run unattended from task start to a CI-ready pull request without intermediate human interaction. (stripe.dev) The Minions harness is a heavily modified fork of Block’s open-source agent Goose, adapted to Stripe’s internal LLM infrastructure and unattended workflows. (infoq.com) Each Minion executes inside a standardized “devbox” — an AWS EC2 instance preloaded with Stripe’s source and services, provisioned from a warm pool in under 10 seconds, and given full shell permissions for isolated runs. (analyticsindiamag.com) For dynamic context, Minions query Toolshed, Stripe’s centralized MCP server, which exposes nearly 500 internal and third‑party tools and supplies each agent with a curated subset relevant to its task. (analyticsindiamag.com) Work is driven by blueprints that interleave deterministic nodes (for linters, tests, and fixed steps) with open-ended agent nodes; Stripe’s standard blueprint permits at most two CI rounds before returning the branch to a human reviewer. (infoq.com) Stripe’s backend is primarily Ruby (with Sorbet typing) and the codebase moves more than $1 trillion in payment volume annually, which Stripe cites as a core reason for building a bespoke, tightly controlled agent stack. (stripe.dev) Engineers demonstrated non‑engineer use and a machine‑to‑machine payments demo where agents autonomously arranged a $5.47 birthday purchase, illustrating both internal adoption beyond engineering and the company’s experiments with agent-initiated transactions. (lennysnewsletter.com)
