How Stripe's homegrown Minions agents turn Slack requests into merged PRs
Stripe built unattended coding agents, called Minions, on a fork of Block's goose agent. They take a task from Slack or an internal tool and return a CI-checked pull request for human review.
The problem
Stripe's codebase runs to hundreds of millions of lines, mostly Ruby with Sorbet and many in-house libraries that LLMs don't know. Developer attention was a scarce resource, and engineers wanted to run many small tasks in parallel.
How they did it
- Engineers start a Minion from a Slack thread, a CLI, a web UI or internal tools (for example, a flaky-test ticket).
- Each run gets an isolated, pre-warmed devbox with no internet or production access.
- The agent loop interleaves model steps with fixed steps for git, linters and tests, and reads the same rule files as Cursor and Claude Code.
- Context comes through an internal MCP server, Toolshed, which has more than 400 tools.
- Local lints run on each push, followed by at most two rounds of CI, before the PR is handed to a human reviewer.
Results as reported by Stripe
- As reported by Stripe: 'over a thousand pull requests merged each week' are produced entirely by Minions and reviewed by humans.
These figures are as reported by Stripe in Stripe engineering blog ↗ (Feb 9, 2026). AgentGid did not measure or independently verify them.
This suits platform or infra teams with a mature CI, linter and sandbox setup, since the setup depends on isolated devboxes and an internal MCP server with 400+ tools, which most teams would need to build. The thousand-PRs-a-week figure is Stripe-reported, and human review effort scales with it. Goose is free (OSS) plus model costs; Pi, also Free (OSS), is a lighter alternative.
The agent used here
Similar use cases
Layer hallucination checks into AI agent workflows with n8n
An n8n blog guide explains how to combine several hallucination detection methods in AI agent workflows and route flagged outputs to retries, error workflows or human review. It s…
The article gives no quantitative results; it is a guidance piece, not a measured deployment. (as reported by n8n)
Doctolib got 47% more pull requests per engineer with Claude Code
Doctolib, the European healthcare booking and practice-software platform, rolled Claude Code out to its whole development team and then built its own remote coding agents on Claud…
The average engineer handled 47% more pull requests between January and September 2026, as reported by Anthropic.
Wayfair rolled out Devin to 1,600+ engineers and merged 40,000+ PRs
Wayfair, the home-goods retailer, moved its engineers onto Devin cloud agents after a bake-off of several platforms. A pilot planned for 20 engineers grew to 100 people and then t…
Wayfair planned for 500 engineers; more than 1,600 now use Devin and have merged more than 40,000 PRs with it, as reported by Cognition.
Guides
Best Claude Code Alternatives in 2026: Open-Source, Cheaper and Multi-Model
The strongest alternatives to Claude Code, grouped by why you are switching: price, model choice, open source or a different workflow. Live usage and prices for each.
What Is MCP (Model Context Protocol)? A Plain Guide and the Most Popular MCP Servers
What the Model Context Protocol is, how MCP servers and clients work, which AI agents support it, how to install a server safely, and the most-starred MCP servers by category.