Stripe Details 'Minions' AI Coding Agents
Engineers at Stripe have detailed their work on internal AI agents called "Minions" that can autonomously handle end-to-end coding tasks. These agents can take a high-level prompt, such as "add a webhook for payment success," and then write, test, and deploy the necessary code. The project points toward a future of "agentic" development where AI can scaffold and maintain significant portions of a company's API layer.
- At Stripe, over 1,300 pull requests are merged each week that are entirely produced by Minions, with the code being reviewed by human engineers before deployment. - An engineer can initiate a Minion's task directly from a Slack thread; the agent then uses the conversation and any included links as its initial context for the coding problem. - The agents operate within isolated cloud developer environments called "devboxes," which are the same sandboxed environments used by Stripe's human engineers, preventing access to real user data or production systems. - The core of the Minion agent is a fork of "goose," an open-source coding agent originally developed by Block. - To gather context, Minions utilize a central server called "Toolshed" that hosts over 400 tools accessible via the Model Context Protocol (MCP), allowing them to query internal documentation, check build statuses, and more. - The workflow, managed by a system called "blueprints," interleaves deterministic code for required steps like running linters with flexible AI agent loops for creative problem-solving. - To prevent endless loops and control costs, a Minion is typically limited to a maximum of two attempts to fix continuous integration (CI) test failures before the task is returned to a human engineer. - Stripe's large and unique codebase, primarily written in Ruby with Sorbet typing rather than a more common framework like Rails, was a key motivation for building a custom, in-house agent.
