Minions

Stripe merges more than 1,000 agent-written pull requests a week through Minions, a figure Alistair Gray revised upward to 1300+ in the second part of his account (2026-02-09, then 2026-02-19). No human writes the code: an engineer describes the task in Slack, the Minion works alone in an isolated cloud environment, and a person reviews the pull request that comes back.

Most of Gray's detail goes to containment. Minions run in devboxes, ephemeral AWS EC2 instances treated as cattle rather than pets, pre-warmed so a full environment is ready in about 10 seconds. The agent reaches neither production nor user data, so Stripe removed the confirmation prompts from its custom fork of goose, Block's open source harness. Orchestration runs through blueprints, a state machine mixing deterministic nodes (clone a repo, cut a branch) with agent nodes for the steps that need judgment. Gray credits that split with saving tokens and improving reliability over a purely agentic approach.

The budget is deliberately small: pre-push lint hooks in under a second, then a maximum of two CI cycles against a suite of more than 3 million tests. Tools arrive through Toolshed, an MCP server exposing around 500 of them, with each agent given a curated subset. Rule files are standardized on the Cursor format and synced to the Claude Code format, so the same context serves Minions and the engineers working alongside them.

Bill Staples reads Stripe's experiments as evidence that the constraint moves next to environment, CI, review and governance.