Stripe has internally developed fully autonomous coding agents called "Minions", capable of producing complete pull requests without human intervention in writing the code. More than 1,000 PRs generated by these agents are merged every week, after human review only.
The workflow is remarkably simple: an engineer sends a Slack message describing the task, the Minion works autonomously in an isolated cloud environment (devbox), produces a PR with passing CI, then a human performs the review. Engineers can launch several Minions in parallel, thereby multiplying their production capacity.
The infrastructure rests on several pillars. Stripe's codebase represents hundreds of millions of lines, primarily in Ruby with the Sorbet type system. Devboxes, pre-warmed cloud environments available in about 10 seconds, provide each agent with an isolated, fully equipped workspace. The agent itself is a custom fork of goose, the open source coding tool created by Block.
MCP integration plays a central role via "Toolshed", a centralized platform hosting more than 400 MCP tools. This architecture gives agents unified access to Stripe's internal systems, letting them interact with the entire development ecosystem without specific configuration.
The CI strategy is particularly deliberate given a test suite of more than 3 million tests. Rather than running everything, the system intelligently selects the tests relevant to each change, with a maximum of 2 CI cycles allowed per Minion. This constraint forces the agents to produce correct code from the first attempts.
The "shift feedback left" philosophy is a guiding principle: lint rules that would fail CI are also enforced in IDEs and git hooks. This way, agents (like human developers) receive feedback as early as possible in the development cycle, reducing costly iterations.
Consistency is ensured by sharing the same rule files between Minions, Cursor and Claude Code. This approach guarantees that coding conventions are followed uniformly, regardless of the tool or agent used.
The Stripe model illustrates a mature vision of AI-assisted development: the human no longer writes code, they specify and validate. Agents become the producers of code, while engineers focus on technical direction, review and quality assurance.